The rapid advancement of embodied intelligence is reshaping the trajectory of intelligent robotics, moving systems away from rigid, pre-programmed execution toward autonomous interaction, environmental adaptation, and complex task completion in physical space. Unlike disembodied artificial intelligence that operates purely in virtual environments, embodied intelligence requires a physical entity to perceive, reason, and act within the real world. Among the many hardware embodiments of this paradigm, the wheeled humanoid robot has emerged as a particularly promising platform. By combining the high mobility of a wheeled chassis with the collaborative manipulation capabilities of dual-arm structures, wheeled humanoid robots offer a compelling balance of operational flexibility, engineering practicality, and deployment readiness. This synthesis of mobility and manipulation positions them as strong candidates for industrial manufacturing, commercial services, household assistance, and special-environment operations.
Recent research and industrial developments increasingly highlight embodied intelligence as the central driver of next-generation robotic systems. Wheeled humanoid robots are no longer viewed as simple mobile manipulators. Instead, they are being designed as integrated platforms in which perception, planning, control, and learning are tightly coupled. This integration allows them to perform long-horizon tasks, adapt to unstructured environments, and collaborate with humans in ways that traditional fixed-base robots or single-arm mobile platforms cannot easily achieve. The following analysis examines the current status, key core technologies, engineering challenges, and future trends of wheeled humanoid robots under the broader umbrella of embodied intelligence.
- Global and Domestic Development Landscape
The international development of wheeled humanoid robots has largely been driven by university research teams and a limited number of start-up companies. Early work focused on the simple combination of a mobile base and one or more robotic arms. Over time, the field has evolved toward tightly coupled architectures that integrate mobility, manipulation, and perception into a unified system. One of the earliest systematic platforms was the Two-arm Mobile Manipulator, or TOMM, developed at the Technical University of Munich. This platform investigated unified modeling, whole-body motion planning, and stability control for a wheeled chassis with dual arms, establishing an important reference for later research on mobile manipulation in embodied intelligence.
Another representative system is Rollin’ Justin from the German Aerospace Center. This platform combines a wheeled mobile base with humanoid dual-arm upper limbs and targets autonomous mobile manipulation in complex environments. Its research emphasis includes whole-body coordinated control, hybrid force-position control, and safe mobile manipulation under constraints. More recently, the team led by Professor Zhang Jianwei at the University of Hamburg has demonstrated significant strengths in multimodal perception fusion and embodied cognitive learning. Their work on dual-arm coordination and whole-body motion generation includes heuristic inverse kinematics frameworks based on multi-objective cost functions, such as BioIK, which effectively addresses high-redundancy dual-arm coordination and collision avoidance. These methods have been validated on wheeled humanoid platforms for complex dual-arm handover and fine sorting tasks. Unlike purely model-driven control, this research emphasizes complementary multi-sensor information and sim-to-real transfer learning to achieve efficient coupling between the wheeled chassis and the manipulators, providing a cross-modal learning paradigm for robust interaction in complex scenarios.
In the industrial sector, the HMND 01 Alpha wheeled dual-arm robot from the British start-up Humanoid represents a recent attempt to build a general-purpose mobile manipulation platform under the embodied intelligence trend. Such platforms emphasize dual-arm collaboration and multimodal perception fusion, targeting multi-task execution and general manipulation capability validation. However, they remain largely in the prototype demonstration and technical verification stage. Overall, international research tends to focus more on fundamental methods and system architecture innovation, while mature commercial products and large-scale applications remain relatively limited. This reflects both the high technical barriers of integrating wheeled mobile dual-arm systems and the expectation that embodied intelligence will accelerate the transition from laboratory prototypes to industrial applications.
In contrast, the Chinese wheeled humanoid robot sector has developed rapidly in recent years, showing a parallel advancement of technical research and engineered products. Chinese efforts place stronger emphasis on system integration, stability assurance, and construction of complex manipulation capabilities oriented toward real-world scenarios, laying a foundation for subsequent application deployment. Several innovative companies have introduced representative wheeled humanoid robots. For example, UBTECH’s full-size general-purpose wheeled humanoid robot Cruzr S2 is equipped with a fourth-generation dexterous hand for sub-millimeter precision operations and integrates learning-based motion control technology to achieve adaptive human-robot collaboration safety in complex environments. Paxini’s TORA DOUBLE ONE is designed specifically for dual-arm collaboration, force perception, and complex interactive manipulation. It adopts a dual steering-wheel AGV chassis with a foldable structure and a dual-modal dexterous hand with nearly 2,000 tactile sensing units, suitable for logistics and warehousing scenarios. Galbot’s GALBOT G1 focuses on general mobile manipulation tasks, emphasizing dual-arm collaboration and autonomous perception. It carries a cross-embodiment navigation foundation model and performs autonomous operations such as material sorting in industrial manufacturing scenarios. The Tencent Robotics X laboratory’s robot “Xiao Wu” features a four-legged wheel-foot hybrid design, large-area tactile skin, multi-fingered dexterous hands, and safe physical human-robot interaction. AUBO’s embodied dual-arm robot emphasizes task execution within an embodied intelligence framework based on a self-developed operating system, highlighting unified mobility, manipulation, and perception. The Tsinghua University spin-off Galaxea’s R1 Lite wheeled humanoid robot is equipped with 23 degrees of freedom, a 360-degree lidar, and multimodal sensors, providing a versatile platform for domestic and international embodied intelligence developers. Youibot strengthens autonomous navigation and dual-arm collaborative operation capabilities in industrial scenarios, forming a complete technology chain covering structural design, system integration, and complex environment adaptation.
Chinese universities have also laid a solid technical foundation in wheeled humanoid robotics. Leveraging intelligent robot control and embedded mechatronics, they achieve real-time multi-sensor fusion and high-precision motion control in system integration. Research also focuses on structural topology optimization and dynamic balance algorithms to improve adaptability in unstructured environments. For instance, the embodied intelligence robotics institute at Beihang University has independently developed a wheeled humanoid robot platform capable of tasks such as heavy object handling, supporting embodied intelligence teaching and research. The “Navigator NAVIAI-WA1” wheeled humanoid robot developed by Professor Xiong Rong’s team at the Zhejiang University Humanoid Robot Innovation Center specializes in ultra-precision operations at the 0.03 mm level and has been applied in production lines of leading enterprises such as Huawei and Geely. Domestic enterprises and universities are gradually maturing in system structure design, sensor configuration, and control architecture, reflecting the rapid iteration capability of the Chinese wheeled humanoid robot sector.
From an industrial application perspective, Chinese wheeled humanoid robots have begun pilot applications in industrial manufacturing, commercial services, and special operations, covering collaborative assembly, material handling, tool operation, and human-robot collaboration. Table 1 summarizes representative Chinese wheeled humanoid robot products and their key specifications. Overall, a preliminary industrial chain covering key components, system integration, and application validation has formed, although further improvements are needed in general task generalization, robustness of whole-body coordinated control, and safety standard systems.
Company Release Time Product Physical and Motion Indicators Perception and Operation Capabilities Functions and Application Scenarios Galbot June 2024 Galbot G1 173 cm / 85 kg / 43 DOF; omnidirectional wheeled leg design with 360-degree movement Adaptable to different terrains; generalizable grasping and environment adaptation Industrial manufacturing, logistics, retail services Tencent September 2024 Xiao Wu 140–180 cm / 80 kg / 29 DOF; multimodal wheel-foot design (dual-wheel, four-wheel, wheel-foot); dual-arm carrying load 50 kg; supports autonomous folding Complex terrain mobility and heavy-load transportation Elderly care, human-robot interaction, special operations Paxini August 2025 TORA DOUBLE ONE 167–169 cm / 70–73 kg / 20–22 DOF; dual-wheel AGV or four-wheel foldable chassis; nearly 2,000 tactile sensing units; hand load 6.5 kg Fine tactile operation Logistics and warehousing, complex interactive manipulation UBTECH August 2025 Cruzr S2 176 cm / 185 kg / 44 DOF; wheeled chassis at 2 m/s; carrying load 15 kg; full-space handling from 0 to 1.8 m Guide, medical transport, production assembly Industrial manufacturing, commercial services, healthcare Unitree November 2025 G1-D Flagship 126–168 cm / 80 kg / 19 DOF; wheeled chassis up to 1.5 m/s; single-arm load 3 kg; battery life 6 h; NVIDIA Jetson Orin NX 100 TOPS Collaborative dual-arm tasks, complex environment mobility Research, education, light industrial tasks RealMan August 2025 RMC-AIDAL 170 cm / 90 kg / 12 DOF; binocular D435 vision plus RGB monitoring; two RM65-B-V ultra-light arms Collaborative dual-arm tasks Research, service, light manufacturing AgiBot October 2025 Genie G2 185 kg / 50 DOF with 26 active joints; high-performance motion joints; high-precision torque sensors Multimodal language interaction Industry, logistics, guidance Star Dynamics June 2025 Q5 165 cm / 65 kg / 44 DOF; compact wheeled chassis; lidar plus vision fusion navigation; single-arm load 10 kg Autonomous obstacle avoidance in narrow passages Commercial services, logistics, inspection Dobot March 2025 ATOM-W 153 cm / 62 kg / 41 DOF; wheeled chassis at 1.5 m/s; repeat positioning accuracy ±0.05 mm; 7-DOF industrial collaborative arms Fine operation such as wafer transfer Electronics manufacturing, precision assembly Keenon March 2025 XMAN-R1 Humanoid structure; 360-degree high-precision perception; 11 multimodal sensors plus large language model Service interaction loop including ordering, food delivery, and table clearing Restaurant and service industry DexForce July 2025 W1 Pro 34 power units; highly humanoid structure; pure vision spatial intelligence sensors Precise control and real-time perception Coffee making, interactive tasks Seer October 2025 X1-PRO 40 DOF; omnidirectional chassis; 45-degree bending structure; single-arm 6 kg, dual-arm instantaneous 16 kg Logistics sorting, loading and unloading, collaborative handling Logistics, manufacturing, warehousing Youibot August 2025 Xunxiao Industrial-grade mobile chassis; “one brain, multiple forms” industrial embodied intelligence model Semiconductor precision manufacturing, power inspection, disaster prevention Special operations, industrial inspection - Key Core Technologies
Wheeled humanoid robot technology integrates mechanical design, sensing, motion planning, intelligent control, and artificial intelligence. It enables robots to move from single-function execution toward adaptive operation in complex environments. The technology can be divided into five major modules: onboard hardware and supporting systems, mobile chassis environment perception and motion planning, dual-arm task understanding and dexterous manipulation, data collection and simulation training for complex scenarios, and embodied intelligence control paradigms with sim-to-real transfer. These modules jointly support mobile manipulation capabilities in diverse scenarios and drive robots from traditional programmed execution tools toward intelligent agents with autonomous perception, decision-making, and learning abilities.
- Hardware Components and Core Architecture
The onboard design of a wheeled humanoid robot is the foundation of its motion and manipulation functions. Its core hardware architecture is typically modular and heterogeneous, mainly including the mobile chassis, lifting and adjustment mechanisms, dual arms, dexterous hands, and end effectors, as well as supporting systems such as motion control, communication, operating systems, and middleware. Current onboard design has formed a mainstream direction of modularization and lightweight construction, but unified design theories and methods are still lacking for whole-machine dynamics optimization under mobile-manipulation coupling and for reliable integration of multi-sensor systems.
In the physical form of a wheeled humanoid robot, the mobile chassis and lifting mechanism together determine the robot’s operating radius and vertical workspace. Mechanical structure selection must balance flexibility, load capacity, and environmental adaptability. For mobile chassis, current mainstream solutions show a trend from simple constraints toward omnidirectional and highly dynamic configurations. Common mobile chassis types include Mecanum wheels, steering wheels, differential wheels, and omnidirectional wheels. Each offers differentiated configurations in omnidirectional movement capability, load capacity, control complexity, and environmental adaptability. To further address the large vertical span of dual-arm workspaces, lifting mechanisms serve as important components connecting the chassis and the waist, enabling the robot to raise and lower its body. Common linear guide rail lifting mechanisms have good rigidity and simple control but are limited by slider stroke and have a small compression ratio. Scissor-folding mechanisms are more compact and can achieve a higher lifting ratio, but they place higher demands on manufacturing precision and lateral force resistance. This combination of mobility and lifting allows wheeled humanoid robots to balance efficient movement on indoor flat floors with flexible reach in complex height environments.
The arm configuration design of wheeled humanoid robots centers on balancing motion flexibility, structural stiffness, and spatial adaptability. Common configurations include serial, parallel, and serial-parallel hybrid types. Among them, serial-parallel hybrid configurations combine the large workspace advantage of serial mechanisms with the high stiffness of parallel mechanisms, becoming an important direction that balances operating range and load capacity. Modular design is a key approach to configuration optimization. By combining standardized joint and link units, it enables flexible structural expansion and convenient maintenance. Combined with lightweight materials, it reduces energy consumption and improves motion response speed. The low-level control of the manipulator focuses on precision control and compliance adjustment. Low-level control methods mainly include position control, force/torque control, and impedance control. Through closed-loop feedback mechanisms, they correct joint motion errors in real time and ensure trajectory tracking accuracy. The application of force control and impedance control technology gives the manipulator adaptive adjustment capability when interacting with the environment or humans, dynamically adjusting motion parameters by sensing changes in contact force to achieve compliant operation.
As the direct carrier of physical interaction between the manipulator and the work object, the end effector design focuses on dexterity, versatility, and sensing integration capability. Humanoid multi-fingered dexterous hands, with anthropomorphic structural design and multidimensional manipulation capabilities, have become the core carrier for complex operations. Dexterous hands adopt multi-finger, multi-joint bionic structural design. Through the combination of flexible joints, tendon ropes, and gear hybrid transmission mechanisms, they can adapt to work objects of different shapes, materials, and stiffness, balancing grasp stability and operational safety. Deep integration of perception and control technology is a core feature that distinguishes dexterous hands from traditional end effectors. By densely integrating tactile, force, and position multimodal sensors at fingertips, finger pads, and palm surfaces, they collect contact pressure, contact pose, and surface morphology information in real time, providing high-precision, high-real-time feedback for low-level control. Their control logic is deeply coordinated with the low-level control of the manipulator. Through real-time interaction of multi-channel signals, they achieve fine adjustment of grasping force and smooth transition of actions, ensuring precise and controllable complex tasks such as precision assembly and in-hand manipulation. Current technical development directions for dexterous hands focus on improving perception accuracy, strengthening multi-task adaptability, and optimizing whole-body collaborative control with manipulators and mobile chassis, providing core support for wheeled humanoid robots to complete complex operations in unstructured scenarios.
Chinese dexterous hand technology has developed rapidly in recent years. The multi-fingered bionic dexterous hand launched by Lingxin Qiaoshou uses a tendon-driven structure and a multimodal tactile sensing system, enabling high-DOF grasping and in-hand manipulation with good application potential in industrial assembly and service robotics. In addition, the Inspire Hand series developed by Inspiring Robots adopts modular design and high-precision servo drive, with multi-finger collaborative grasping capability, and has been widely used in research and education. Continuous optimization in structural design, tactile perception, and control algorithms provides important technical support for wheeled humanoid robots to achieve complex manipulation tasks.
As a key infrastructure connecting the robot’s mechanical body and upper-level intelligent algorithms, the hardware structure and software architecture of a wheeled humanoid robot together form its core computing and control system. On one hand, the system requires a high-computing-power platform to support multimodal perception processing and intelligent decision-making. On the other hand, it must meet the strict real-time requirements of chassis motion control and manipulator servo drive. Therefore, modern wheeled humanoid robots generally adopt a layered architecture with coordinated design of hardware computing architecture and software system platform to achieve efficient collaborative operation of perception, planning, and control tasks. For complex control chains and multi-source data processing requirements, research and engineering practice typically adopt a layered heterogeneous computing architecture and use a unified software platform to schedule heterogeneous hardware resources and abstract functions.
In terms of hardware computing architecture, to balance the computing power requirements of high-level intelligent algorithms and the real-time requirements of low-level motion control, the system usually adopts a “brain-cerebellum” layered computing architecture. The “brain” part carries a high-performance CPU or GPU computing platform for high-level intelligent algorithms such as environmental perception, semantic understanding, task planning, and vision-language models. The “cerebellum” part consists of real-time embedded controllers or industrial controllers, responsible for real-time tasks such as chassis motion control, joint servo drive, and high-frequency sensor data acquisition. Through this heterogeneous computing structure, millisecond-level control response can be guaranteed while providing sufficient computing resources for complex intelligent algorithms, thereby achieving collaborative operation of mobility, manipulation, and perception tasks. In terms of software system architecture, robot operating systems and middleware technologies constitute an important foundation connecting hardware devices and upper-level applications. Current wheeled humanoid robots generally adopt a dual-system architecture with a general-purpose operating system and a real-time operating system. The general-purpose operating system handles upper-level algorithm execution, data processing, and software ecosystem support, while the real-time operating system focuses on low-level motion control and real-time task scheduling. On this basis, robot middleware represented by the Robot Operating System (ROS) enables data interaction and functional collaboration among multiple modules through a node-based communication mechanism. Its publish-subscribe communication mechanism can build high-bandwidth, low-latency data exchange channels, greatly reducing the development complexity of multi-sensor fusion, dual-arm collaborative control, and task scheduling systems. Meanwhile, the rich functional packages and development tools in the ROS ecosystem enable rapid integration of SLAM, motion planning, and visual processing modules into the robot system.
In low-level device communication, industrial fieldbuses and real-time communication protocols provide key support for stable hardware operation. Industrial communication buses such as EtherCAT and CAN can achieve unified addressing and high-speed communication for chassis drivers, joint servo systems, and various sensor devices, and realize system-level time synchronization through distributed clock mechanisms, ensuring coordinated operation of multiple actuators in complex tasks. In addition, to reduce system integration complexity and improve software scalability, modern robot systems usually build a Hardware Abstraction Layer (HAL) to uniformly encapsulate low-level hardware resources. This layer shields differences among hardware platforms, communication protocols, and mechanical structures through standardized interfaces, and uniformly encapsulates chassis mobility, manipulator operation, and sensing system capabilities into standardized software interfaces or functional modules, enabling upper-level algorithms to be developed and deployed in a unified software environment. This system design concept centered on software architecture gives the robot system good modular expansion capability, facilitating rapid integration of heterogeneous hardware platforms and providing a stable foundation for deployment and iteration of complex intelligent algorithms.
- Mobile Chassis Environment Perception and Motion Planning
Environmental perception and motion planning for the mobile chassis are the foundation for autonomous mobile operation of wheeled humanoid robots. The core issues include environment modeling, pose estimation, path generation, and trajectory tracking. Compared with traditional indoor mobile robots, wheeled humanoid robots face more dynamic and open application environments, imposing higher requirements on perception robustness, planning real-time performance, and control accuracy. Existing mobile chassis navigation methods can be summarized from two levels: path planning strategy and control implementation. They mainly include graph search planning, sampling planning, local planning, optimization control, and recently developed learning-based navigation and semantic navigation methods. Table 2 summarizes typical methods for mobile chassis environment perception and motion planning.
Method Category Representative Algorithms Advantages Limitations Typical Applications Graph search planning A*, D Lite Good stability, strong interpretability Weaker adaptability to dynamic environments Global path planning Sampling planning RRT, RRT*, PRM Suitable for complex space search Path smoothness is average Feasible path generation Local planning DWA, TEB Strong real-time performance Prone to local optima Dynamic obstacle avoidance Optimization control MPC Can handle dynamic constraints Large computational load Trajectory tracking control Learning-based navigation Deep reinforcement learning Strong adaptability in complex environments High training cost, insufficient stability Unknown environment navigation Semantic navigation Semantic Navigation Supports target-level navigation Depends on semantic perception accuracy Service scenario navigation In environmental perception and localization mapping, a single sensor is difficult to simultaneously guarantee accuracy, robustness, and adaptability. Current mainstream solutions mostly adopt a multi-sensor fusion architecture of lidar, vision, and inertial measurement units (IMU). Among them, lidar is suitable for obtaining stable geometric structure information, vision sensors can provide texture and semantic information, and IMU can enhance short-term motion estimation stability. Through tightly coupled fusion of multi-source information, the mobile chassis can achieve more stable localization and mapping, and gradually evolve from geometric maps to map representations that integrate scene semantic information.
In path planning and motion control, existing methods usually adopt a hierarchical structure combining global planning and local planning. Global planning generates reachable paths based on a prior map, commonly using graph search algorithms such as A* and D Lite, as well as sampling planning algorithms such as RRT, RRT*, and PRM. Local planning targets dynamic obstacles and local passage constraints, often using the Dynamic Window Approach (DWA) and the Timed Elastic Band (TEB) method for real-time obstacle avoidance. Graph search methods have good stability, sampling planning is more suitable for complex spaces, and local planning has strong real-time performance in dynamic scenarios. At the trajectory execution level, the mobile chassis needs to achieve stable tracking of the planned path through closed-loop control. Among these, Model Predictive Control (MPC) can explicitly consider kinematic and dynamic constraints within the prediction horizon, so it is widely used in mobile chassis trajectory tracking control. On this basis, research has further introduced semantic navigation methods to assist path decision-making by identifying semantic targets such as doors, tabletops, and passages, thereby improving navigation efficiency in service and indoor operation scenarios. In addition, navigation methods based on deep reinforcement learning have gradually been used for autonomous obstacle avoidance and strategy learning in unknown environments, showing certain adaptability in complex scenarios, but their engineering stability and training cost still require further study.
- Task Understanding and Dual-Arm Dexterous Manipulation
Compared with single-arm manipulation, the core problem faced by wheeled humanoid robots is no longer a single grasp or trajectory generation, but task-centered dual-arm collaborative understanding and execution. Dual-arm tasks usually require clear manipulation objectives, constraints, and execution sequences, including the geometric and functional attributes of target objects, safety and collision avoidance constraints during operation, and the temporal relationships of multi-step tasks. On this basis, research generally divides dual-arm collaboration modes into several typical categories: master-slave collaboration, where one arm performs the main operation and the other assists with support or constraint; symmetric collaboration, which emphasizes equivalence in load distribution and motion execution between the arms, often used for collaborative carrying and symmetric assembly; and hybrid collaboration, which dynamically switches arm roles according to task stages, closer to real complex task requirements. This task representation provides high-level semantic constraints for subsequent motion planning and control.
In dual-arm collaborative motion planning, existing research centers on basic planning methods and mainly includes three technical routes. The first is inverse kinematics-based planning, which jointly solves dual-arm joint configurations to satisfy end-effector poses and collaboration constraints, featuring high computational efficiency and easy integration. The second is sampling-based planning, such as introducing dual-arm coupling constraints in a high-dimensional configuration space to improve the ability to search for feasible solutions in complex environments. The third is optimization or task-space constraint representation methods, which explicitly model grasp pose families, tool coordinate frames, and constrained motion trajectories as optimization objectives or constraints, more suitable for fine assembly and tool operation tasks. These methods constitute the technical foundation of current dual-arm collaborative planning.
At the manipulation-side perception and system integration level, dual-arm dexterous manipulation places stringent requirements on the dimensionality and precision of multimodal information fusion. A representative example in such multimodal fusion and frontier dexterous manipulation exploration is the RoboFusion robot body interconnection architecture jointly launched by the Tencent Robotics X laboratory and the Futian laboratory. This architecture builds a full-stack communication system from the physical layer to the application layer, supports high-precision clock synchronization and real-time data transmission, and can seamlessly integrate into mainstream robot development ecosystems, enabling advanced functions such as multi-joint collaboration and efficient OTA upgrades. In addition, the OPENARM dual-arm platform led by Enactic and other technology teams provides an open-source hardware demonstration for contact-intensive physical AI research. Its onboard design balances high dynamic response and compliant safety, accelerating the validation loop of VLA large models, offline reinforcement learning, and sim-to-real strategies in complex dual-arm collaborative tasks.
- Data Collection and Simulation Training for Complex Scenarios
High-quality data is the cornerstone of embodied intelligence evolution, and teleoperation technology is a key pathway for obtaining high-fidelity physical interaction data. For long-duration and contact-intensive tasks, the ALOHA system developed by Stanford uses low-cost master-slave isomorphic mapping. Compared with VR solutions lacking force feedback, it effectively reduces the complexity of inverse kinematics solving and improves operation accuracy. To break through laboratory scenario limitations, the UMI system proposed by Chi et al. combines a handheld fisheye camera with SLAM technology to achieve portable in-the-wild data collection. With the development of mobile manipulation, teleoperation is evolving toward whole-body coordination of wheeled humanoid robots. The Mobile ALOHA designed by Fu et al. achieves collaborative input of the chassis and dual arms through whole-body teleoperation, solving the safety control problem under nonholonomic constraints. While ensuring low latency and high presence, it provides an efficient technical means for interactive data collection in complex dynamic environments. Table 3 summarizes the technical comparison of teleoperation and data collection.
Technology Category Core Features Advantages Limitations Core Applicable Scenarios VR teleoperation Head-mounted visual immersion plus 6-DOF controller pose tracking Intuitive spatial perception, safe and uncoupled operation, easy access to consumer-grade hardware Lack of force feedback, high difficulty in human-robot motion retargeting, prone to dizziness during long operations Long-range navigation, non-contact pick-and-place tasks Motion capture Optical markers or pure vision contactless human motion tracking Natural and unconstrained operation, supports whole-body motion capture, adaptable to complex dexterous gestures Susceptible to occlusion, precision fluctuation in pure vision solutions, inherent errors in human-robot mapping Dexterous hand fine manipulation, complex gesture interaction tasks Exoskeleton/master-slave arm Mechanical coupling and master-slave isomorphic mapping High acquisition precision, native force feedback, data can be used directly without complex retargeting High hardware cost, limited operating space, high physical consumption during long operations Precision assembly, contact-intensive manipulation tasks Force feedback device Desktop pen-type force interaction device High force rendering precision, excellent micro-manipulation control performance Only supports end-point control, limited workspace, steep operation learning curve Remote surgery, micro-manipulation, rigid object surface exploration Kinesthetic teaching Directly dragging the robot end effector or joints for teaching Zero additional hardware cost, no pose mapping error Force sensor data susceptible to interference, prone to visual occlusion, only suitable for light collaborative arms Simple trajectory reproduction, trajectory planning tasks without visual input To maximize data value, it is necessary to establish strict time-synchronized acquisition specifications including vision, force perception, joint states, and odometry. The core value of data lies not only in successful demonstrations but also in recording failures and recovery processes. Google DeepMind confirmed in RT-1 that closed-loop data containing anomaly recovery can significantly improve model robustness. In addition, Mandlekar et al. pointed out that segmenting, cleaning, and detecting contact events in long-chain trajectories can effectively improve the efficiency of offline reinforcement learning. The Open X-Embodiment project further demonstrated that generalizable data with task diversity, environment diversity, and coverage of various disturbances, combined with unified action space annotation, is key to achieving cross-scenario and cross-platform skill transfer.
To address the scarcity of physical data, high-fidelity simulation based on physics engines has become a mainstream solution. The Isaac Gym proposed by Makoviychuk et al. uses GPU parallel computing to greatly improve sampling efficiency and supports large-scale domain randomization to cover the distribution of physical parameters in the real world. ManiSkill2 developed by Gu et al. enhances skill generalization through fine physical property modeling and multi-sensor simulation. Recently, scene construction technology combined with generative AI has become a new trend. The Holodeck proposed by Yue et al. uses large language models to automatically generate semantically rich 3D scenes, solving the time-consuming problem of traditional simulation scene construction and providing a large-scale training environment with physical realism and visual diversity for mobile manipulation, strongly promoting zero-shot sim-to-real transfer. Table 4 mainly compares the characteristics of mainstream embodied intelligence robot simulation platforms. The current development of simulation platforms shows a core feature of parallel specialization and generalization: platforms for precision manipulation tasks continue to deepen physical interaction fidelity, while platforms for large-scale reinforcement learning focus on computing efficiency and scene generalization capability. The technical boundaries of the two are gradually merging.
Simulation Platform Core Engine Core Features Advantages Limitations Applicable Task Types SAPIEN (base of ManiSkill series) PhysX (customized) Articulated object part-level physical interaction High physical interaction fidelity, complete open-source ecosystem, rich supporting datasets Limited visual rendering capability, insufficient GPU parallel efficiency in early versions Precision manipulation, tool use, part assembly Isaac Sim (NVIDIA) PhysX 5 (Omniverse) GPU-parallelized large-scale simulation Leading training efficiency, realistic visual rendering, well-integrated ROS ecosystem High hardware threshold, closed-source core engine, prone to clipping in tiny contact simulation Motion control, large-scale reinforcement learning, general grasping AI2-THOR (Allen Inst.) Unity Semantic household interactive environment Rich scene semantic information, high visual fidelity, active community ecosystem Simplified physical model, relatively low frame rate Long-horizon task planning, visual navigation Habitat (Meta / FAIR) Bullet (v2/v3) High-frame-rate navigation simulation based on real scene scanning Extremely high single-machine frame rate, scenes built from real house scans Weak object interaction capability, mainly rigid body simulation, no support for soft body or fluid simulation Visual navigation, basic mobile manipulation RFUniverse Unity Multi-physics full-modal simulation Supports fluid, soft body, and gas multi-physics simulation, complete VR interface High computational overhead for multi-physics, high API learning difficulty Complex physical interaction, flexible object manipulation, daily household tasks ThreeDWorld (TDW) Unity Multimodal perception physical simulation Unique physics-based audio rendering capability, flexible procedural scene generation Low popularity in robot manipulation field, non-mainstream interface Audio-visual fusion learning, physical commonsense reasoning - Embodied Intelligence Control Paradigms and Sim-to-Real Transfer
In mobile manipulation tasks facing complex dynamic environments, the core problem of the control system is not only low-level motion control accuracy but also how to achieve a stable closed loop from semantic goals to executable actions under long-horizon tasks, perception uncertainty, and contact interaction constraints. The narrative follows a main line of “control paradigm–intelligent decision-making tool–training and deployment mechanism.” First, two main control paradigms, hierarchical control and end-to-end control, are summarized. Then, the role of multimodal large models as tools in high-level task understanding, semantic planning, and skill scheduling is discussed. Finally, how sim-to-real transfer provides a training, validation, and deployment loop for the above control strategies is analyzed. Among them, the hierarchical control paradigm usually decouples robot tasks into three levels: high-level logical planning, mid-level skill primitives, and low-level motion control, thereby achieving dimensionality reduction and modular solution of control problems. The high-level planning layer generates physically feasible sub-goal sequences for task objectives and environmental states, and can constrain and screen candidate actions through evaluation functions such as operability, safety distance, and energy consumption. The mid-level module builds a skill library covering parameterized primitives such as grasping, pushing, delivering, and opening-closing, achieving semantic connection between task planning and motion control. The low-level control is responsible for high-frequency closed-loop execution of the chassis, manipulators, and end effectors, ensuring trajectory accuracy and contact compliance through position control, force/torque control, or impedance control. In addition, the system can integrate anomaly monitoring and recovery mechanisms based on perception feedback. When target deviation, grasping failure, or collision risk is detected, local replanning is triggered, forming an operation logic with closed-loop self-correction capability.
Unlike hierarchical control, which relies on explicit task decomposition and intermediate skill representation, end-to-end control is another data-driven control paradigm. Its core is to use deep neural networks to establish a direct mapping from raw perception input to action commands. In the field of imitation learning, Diffusion Policy (DP) models action generation as a conditional diffusion process, which can better represent multimodal action distributions and alleviate the covariate shift problem of traditional regression models in complex manipulation tasks. In the field of reinforcement learning, strategies can explore optimal solutions for operations through the principle of cumulative reward maximization, and combine large-scale offline data pre-training with online fine-tuning to improve the generalization ability of visual grasping and mobile manipulation strategies. The advantage of end-to-end methods is that they can implicitly learn the associations among object properties, contact modes, and action sequences from data, but their sample efficiency, interpretability, and safety verifiability remain limited. Therefore, in engineering deployment, they usually need to be combined with high-fidelity simulation, constraint checking, and safety control modules.
Beyond the above control paradigms, multimodal large models are more appropriately positioned as tools for task understanding and high-level decision-making. For example, the PaLM-E framework uses large models mainly as high-level reasoners to convert vague natural language instructions into structured planning prompts rather than directly outputting low-level continuous control signals. The RT-2 framework uses a vision-language-action (VLA) model to discretize robot actions into “tokens” that can be autoregressively predicted by a large model, thereby improving generalization ability in open-vocabulary tasks. Some studies also use LLMs to generate code or build 3D value maps, combining semantic understanding capability with traditional motion planning methods. Therefore, the main role of large models in wheeled humanoid robot systems is to provide task decomposition, objective constraints, and skill scheduling bases for hierarchical control, or to provide semantic conditions and prior knowledge for end-to-end strategies. Considering that large models may have hallucinations, insufficient physical commonsense, and unclear safety boundaries, feasibility verification, collision detection, and safety constraint mechanisms still need to be introduced in actual deployment.
Sim-to-real transfer is a core technical means to bridge the “gap” between simulation environments and real physical scenarios and to reduce the cost of acquiring embodied data. Domain randomization forces policy networks to extract robust features by perturbing physical attributes and visual parameters in the simulation environment over a wide range, and has been successfully applied to complex dexterous hand manipulation tasks. However, offline domain randomization alone is difficult to cover factors such as wheel-ground contact, load disturbances, joint friction, and sensor noise that change over time in real systems. Therefore, between domain randomization and online system identification, a transition step of “parameter adaptation and latent variable estimation” is needed. On one hand, simulation randomization expands the dynamics distribution that the policy encounters during training. On the other hand, the robot uses proprioceptive history, action inputs, and state responses during execution to estimate implicit parameters of the current environment or body dynamics, and uses them as policy condition inputs, achieving a connection from offline robust training to online adaptive control. Within this framework, algorithms such as RMA can be understood as representatives of the online adaptive path. They estimate environmental latent variables such as ground friction, load changes, and dynamic deviations from short-term proprioceptive history through an adaptation module, and then drive the base policy to adjust action output in real time, improving the system’s adaptability to uncertain physical changes.
In summary, the embodied intelligence control technology of wheeled humanoid robots can be understood at three levels: control paradigm, intelligent decision-making tool, and training and deployment mechanism. Hierarchical control and end-to-end control represent two main control paradigms: explicit task decomposition and data-driven mapping. Multimodal large models mainly undertake high-level decision-making functions such as task understanding, semantic reasoning, and skill scheduling, and provide objective constraints and prior knowledge for control strategies. Sim-to-real transfer supports the training, validation, and engineering deployment of control strategies through domain randomization, parameter adaptation, online system identification, and real data回流. These technologies are located at different levels in the system chain of “perception–decision–planning–control–execution” of wheeled humanoid robots. Their collaborative integration is a key direction for improving long-horizon task execution capability, open-scenario generalization ability, and engineering deployment reliability.
- Hardware Components and Core Architecture
- Typical Applications
Wheeled humanoid robots have already completed engineering pilot applications in industrial manufacturing, commercial services, special operations, and other fields. A development pattern of parallel advancement in technology research and engineering implementation has formed, with a preliminary industrial system covering core components, complete machine research and development, and scenario applications. The following sections outline typical application scenarios and their characteristics.
- Industrial Manufacturing Applications
In industrial manufacturing, wheeled humanoid robots are gradually being applied to material handling, precision assembly, and equipment inspection in flexible factories. Compared with fixed-station robots, they have spatial mobility based on a wheeled chassis, enabling cross-regional operations in multi-station and multi-cycle production environments. The dual-arm structure supports complex tasks such as collaborative grasping, alignment assembly, and tool operation. This form has strong adaptability to manufacturing models with fast production cycles, frequent process changes, and small-batch, multi-variety production, helping to alleviate the shortcomings of traditional automated production lines in flexibility and scalability.
In specific applications, new energy vehicles and 3C manufacturing are typical early deployment scenarios. In tasks such as power battery module assembly, wire harness organization, and precision component assembly, dual-arm collaboration can significantly improve operation stability and assembly accuracy. In equipment inspection and maintenance, mobile manipulation capability enables the robot to complete autonomous operations from detection to disposal. Several Chinese enterprises have achieved early industrial deployment of wheeled humanoid robots. UBTECH’s Cruzr S2 wheeled humanoid robot adapts to all types of industrial scenarios and achieves large-scale commercial deployment. Multiple units can be flexibly deployed in core production workshops, significantly improving production line automation and flexible production efficiency. AgiBot’s Expedition A2-W has achieved large-scale industrial deployment, with nearly one hundred units deployed in core workshops, enabling continuous operation and greatly reducing manual intervention costs. Leju’s KUAVO 5-W features ultra-long endurance, ultra-large load, and high efficiency and stability. Combined with a data-driven industrial embodied model and the KUAVO robot industrial software platform, it is highly compatible with industrial production cycles. Galbot S1 is the first deployed 50 kg heavy-load product, targeting high-intensity manual operations such as cross-station transfer and collaborative assembly of heavy materials in industrial scenarios, achieving fully autonomous operation and providing a new technical path for flexible upgrading of heavy-load manufacturing scenarios. Fuwit has deployed precision assembly scenarios in multiple fields, efficiently completing various collaborative tasks.
In special environment operations, such as substation and large facility inspection, dual-arm mobile manipulation systems can perform simple operations or emergency handling while completing inspections. From industrial practice, wheeled humanoid robots are reconstructing the automation model of industrial manufacturing. They break the limitations of traditional fixed-station robots and, with the characteristics of one machine with multiple capabilities and full-area coverage, adapt to the current flexible production needs of small batches, multiple varieties, and rapid iteration, providing a replicable and scalable engineering path for the implementation of embodied intelligence technology in industrial manufacturing.
- Commercial Service Applications
In commercial services, the capabilities of wheeled humanoid robots are gradually expanding from single delivery to a combination of delivery, manipulation, and human-robot interaction. In scenarios such as unmanned delivery, supermarket guidance, restaurant service, and medical rehabilitation, wheeled humanoid robots need not only autonomous navigation but also stable grasping, placement, and human-robot interaction capabilities. The dual-arm structure supports multi-item handling, two-handed delivery, and more natural interactive actions, while emphasizing safety and reliability, thereby improving service efficiency and user experience.
Wheeled humanoid robots in commercial service scenarios are evolving from perception-oriented to manipulation-oriented. For example, Keenon’s first wheeled mobile dual-arm service robot XMAN-R1 covers multiple scenarios such as welcoming, cleaning, and item delivery through natural language interaction and anthropomorphic actions, demonstrating the application prospects of wheeled humanoid robots in commercial services. Galbot uses its self-developed Galbot wheeled humanoid robot to build an embodied intelligence retail solution for normalized consumer operations, achieving fully autonomous operation of product recommendation, ordering and payment, precise grasping, and face-to-face delivery, promoting breakthrough applications in retail service scenarios. Pudu Robotics has released a commercial wheeled mobile dual-arm embodied intelligence service robot, integrating multiple technology stacks such as mobility, manipulation, and interaction, with generalizable manipulation capabilities. The release of these robots marks the official start of commercialization of wheeled mobile dual-arm service robots. Commercial service robots are entering a new era of embodied intelligence, which will drive the industry to upgrade from “single-point function replacement” to “full-scenario service coverage.”
- Household Applications
In household applications, potential uses of wheeled humanoid robots include item organization, life assistance, and elderly care support, such as storage and organization, automatic cleaning, and tool use. These tasks usually have highly unstructured characteristics, imposing higher requirements on environmental understanding, operation stability, and safety. Dual-arm collaboration has obvious advantages in dealing with diverse objects and complex operations. In education, entertainment, and emotional companionship, dual-arm mobile robots can enhance user experience through richer action expressions and interaction methods. Combined with aging-friendly service needs, they have important application prospects in elderly companionship, daily auxiliary operations, and safety monitoring.
For example, Galbot G1 can complete cleaning, orderly organization, accurate item finding, and express delivery retrieval in household scenarios. The AlphaBot 2 wheeled humanoid robot from Zhiyuan Robotics can control home appliances, organize items, and perform high-difficulty operations such as tea making and cooking. The latest generation wheeled humanoid robot Wanda2.0 from Youliqi can provide multi-version software and hardware adaptation solutions according to different customer scenario needs. It can provide emotional companionship and complete delicate and fragile operations such as bartending and dishwashing. The elderly care robot “Xiao Wu” developed by the Tencent Robotics X laboratory has multiple capabilities including safe physical human-robot interaction. It can assist in caring for patients and the elderly and can achieve functions such as lifting and shifting, active suspension, and folding and unfolding. The household scenario is a typical representative of unstructured environments and is one of the ultimate tracks for wheeled humanoid robots to move from industrial and commercial scenarios to the civil market. Although household applications still face challenges such as cost, reliability, and long-term autonomous operation, with the continuous maturity of perception, planning, and embodied intelligence technologies, wheeled humanoid robots are expected to play a more important role in future household services.
- Industrial Manufacturing Applications
- Challenges and Bottlenecks
Although wheeled humanoid robots have made significant progress in onboard design, mobile manipulation collaborative control, and embodied intelligence methods, their stable operation and engineering application in complex open environments still face many constraints. Related problems are not concentrated in a single technical module but run through key links such as hardware integration, environmental perception, collaborative planning, task learning, and safety standards. The main challenges can be summarized as core components and system integration, perception system robustness, collaborative dynamic planning for mobile manipulation, long-horizon complex task generalization, and safety and standardization.
- Core Components and System Integration
The overall performance ceiling of a wheeled humanoid robot is essentially determined by the level of component integration and system architecture design capability. Current contradictions in component selection and system integration include the need for high-performance, high-reliability execution and perception units on one hand, and limitations in cost, mass, volume, and endurance on the other. The superposition of chassis movement, dual-arm manipulation, multimodal sensing, and computing loads makes the peak and average power consumption significantly higher than that of single-arm or single mobile platforms, causing endurance shortages and thermal management problems. Current energy consumption is mainly concentrated in high-torque servo drives, frequent start-stop and acceleration-deceleration motions, end-effector grasping or tool execution, and high-computing-power perception and planning units. Excessive motor margins, heavy mechanical structures, and inefficient motion planning further amplify energy losses. Therefore, optimization cannot remain at the component level alone. It must simultaneously consider high-efficiency motors and drives, lightweight structural materials, regenerative braking and energy management, and collaborative optimization of paths and actions. For heterogeneous wheeled humanoid robots, it is also necessary to establish relatively accurate energy prediction and management models; otherwise, task interruption and scheduling failures are likely in multi-robot collaboration scenarios. Existing research still has the problem of disconnection between component-level optimization and system-level collaboration. Attention to the dynamic variation of overall energy consumption under mobile-manipulation coupling conditions is still insufficient. Future work should build an energy collaborative optimization framework around components, systems, and tasks to alleviate endurance bottlenecks.
To support dual-arm high-DOF manipulation and whole-body collaborative control, wheeled humanoid robots usually need high-bandwidth servo drives, torque and current sensing units, and high-performance graphics processors and AI accelerators as core functional components. However, such components often come with large volume, high heat dissipation and power supply requirements, and complex wiring, creating obvious contradictions with the limited installation space of the mobile chassis and the lightweight requirements of the whole machine. Execution units need to balance dynamic response, impact resistance, and long-term operation stability. Computing units need to support multiple tasks such as perception, planning, and control under limited power consumption. This makes the collaborative integration of drive, sensing, computing power, and power supply a key difficulty in engineering implementation. Existing solutions often sacrifice some performance in exchange for miniaturization and deployability, while still facing significant engineering obstacles in packaging processes, thermal design, and system adaptation. Future efforts need to rely on the development of power semiconductors, heterogeneous integrated chips, and high-density packaging technology to promote efficient collaborative integration of execution, sensing, and computing units.
Dexterous hands and multifunctional end tools can significantly improve the grasping and manipulation capabilities of wheeled humanoid robots, but they also introduce new complexity at the system integration level. Current high-DOF dexterous hands generally suffer from complex structures, dense drive and sensing units, insufficient reliability, and difficult maintenance. Especially in industrial and service scenarios, sensor drift, component wear, and environmental interference can quickly degrade their working performance. Increasing degrees of freedom and multimodal perception capability usually means more drive links and data links, further increasing wiring complexity, signal interference risk, and control computing burden. For wheeled humanoid robots, it is also necessary to arrange end-tool quick-change interfaces, power and signal channels, and necessary safety sensing units in a limited wrist space. Therefore, standardization and modular design of mechanical, electrical, and software interfaces have become key issues in integration.
At the same time, to improve contact operation capability and human-robot collaboration safety, wheeled humanoid robots are gradually introducing new sensing units such as electronic skin, flexible tactile arrays, deformation sensors, distributed force sensors, and multimodal perception networks. However, such sensors still have obvious deficiencies in packaging methods, mechanical fatigue resistance, noise and drift control, and calibration compensation, and their integration compatibility with rigid structures and traditional control hardware is limited. Specifically, flexible tactile arrays based on conductive polymers and liquid metals are prone to conductive network damage and signal drift under large deformation cyclic loading, and precise decoupling of pressure and shear force still needs to be broken through. Distributed force sensing solutions such as fiber Bragg gratings and flexible capacitive arrays face engineering problems such as bending fatigue, local failure, and fault-tolerant reconstruction. In addition, multimodal sensor fusion also involves high-dimensional data real-time processing, communication bandwidth, and latency control, requiring the combination of edge computing, hierarchical fusion architecture, and real-time bus design to achieve efficient collaboration between the sensing system and control hardware. Therefore, the development of end effectors and multimodal sensing systems should not simply pursue improvements in degrees of freedom and sensing density, but should pay more attention to reliability, maintainability, and integration cost in actual scenarios.
From the whole-machine level, the system integration difficulty of wheeled humanoid robots is not limited to whether various components “fit.” It is more about long-term stable operation capability after integration. High-DOF joints, multi-source sensing systems, and complex communication links improve platform capabilities but also significantly increase potential failure points. Under continuous operation conditions, component wear, heat accumulation, loose connections, signal drift, and software anomalies can all lead to performance degradation or even task failure. Compared with short-term validation in the laboratory, engineering applications pay more attention to system stability and maintainability under repeated loads, continuous vibration, and complex interference conditions. In addition, current wheeled humanoid robots still lack a relatively complete standardization foundation in interface design, module replacement, and system reuse. Different platforms differ greatly in drive units, sensor interfaces, end effector mounting methods, and software communication protocols, increasing system development and maintenance costs and limiting platform expansion and industrial collaboration. Overall, the system integration difficulty of wheeled humanoid robots is not a single component performance deficiency but a lack of system-level design methods and standardization systems oriented toward mobile-manipulation integration. This is also a key link that must first be broken through in the transition from laboratory prototypes to large-scale engineering deployment.
- Perception System Robustness
Perception system robustness is an important foundation for wheeled humanoid robots to achieve autonomous operation in open environments and is a prominent difficulty that distinguishes them from traditional fixed manipulators and single mobile platforms. Wheeled humanoid robots not only need to complete environmental perception and autonomous localization but also need to continuously acquire multi-source information such as robot body state, target objects, obstacles, and surrounding personnel state under parallel mobility, manipulation, and interaction. In unstructured scenes, personnel movement and object pose changes can easily cause pre-built maps to fail, leading to localization loss. Therefore, improving the robustness of autonomous localization and mapping systems under drastic environmental changes is key to ensuring continuous execution of mobile manipulation tasks. The current mainstream path in this field is to combine semantic information for dynamic feature removal and use appearance-feature-based fast relocalization technology to improve system recovery capability.
In addition to scene-level perception, wheeled humanoid robots also need to acquire finer state information for the manipulation process. For example, when grasping transparent, reflective, flexible, or stacked objects, single visual information often cannot provide a stable and reliable basis for manipulation. In dual-arm collaborative assembly, constrained manipulation, and tool use processes, relying only on external vision is also difficult to accurately determine contact states and fit relationships. Therefore, collaborative fusion of multimodal information such as vision, laser, inertia, force perception, and touch is a necessary way to improve the perception robustness of wheeled humanoid robots. It should be noted that multimodal perception is not equivalent to the simple superposition of sensors. Different sensors have obvious differences in sampling frequency, noise characteristics, spatiotemporal synchronization, and coordinate representation. Cross-modal fusion also introduces new uncertainty propagation problems. Future research needs to focus on dynamic environment adaptation, long-term operation stability, and multimodal information consistency to provide more reliable state input for subsequent planning and control.
- Collaborative Dynamic Planning for Mobile Manipulation
The strong dynamic coupling between the chassis and dual arms is the essential difficulty of collaborative planning. Unlike traditional separated control modes of “navigate first, then manipulate,” wheeled humanoid robots usually need to simultaneously adjust chassis pose, dual-arm trajectories, and end-effector manipulation strategies during motion to ensure continuity and efficiency of task execution. The inertial disturbance of chassis motion directly affects dual-arm manipulation accuracy, and the reaction force of dual-arm motion disrupts chassis force balance. Traditional separated control modes are difficult to adapt to efficient operation needs. Current collaborative planning faces three core contradictions: the contradiction between modeling accuracy and computational efficiency of global coupled models, the contradiction between trajectory optimality and dynamic adaptability, and the contradiction between large-range operation and whole-machine dynamic balance. Although existing methods such as whole-body model predictive control can partially solve coupled constraint problems, they still face computing bottlenecks and insufficient robustness in highly dynamic scenarios. How to achieve efficient solution of coupled models and real-time trajectory optimization while ensuring system stability is a core obstacle restricting expansion from structured scenarios to unstructured scenarios.
- Generalization for Long-Horizon Complex Tasks
Insufficient generalization capability for complex tasks is a core bottleneck restricting the transition of wheeled humanoid robots from standardized scenarios to open general scenarios. Its challenges run through the entire chain from task semantic understanding to skill transfer learning and physical execution. At the level of autonomous decomposition of long-horizon complex tasks, the core difficulty lies in semantic understanding of abstract instructions and causal logical reasoning. Existing large-model-based task planning methods can complete decomposition of fixed tasks in known scenarios, but in open scenarios with drastic environmental layout changes, unknown object properties, and ambiguous instructions, they cannot stably generate sub-goal sequences that conform to physical rules and precondition logic. The essence is that existing models lack deep implicit understanding of causal relationships in the physical world and cannot achieve cross-scenario generalization of task logic.
At the level of cross-scenario skill transfer, the core constraints come from two dimensions. First, there is a systematic lack of high-quality embodied datasets. Existing data collection costs are high, annotation specifications are not unified, and most focus on single successful demonstrations, lacking coverage of disturbances, failures, and recovery processes, thus failing to provide sufficient generalization training samples for models. Second, there is the inherent gap between simulation and reality. Existing simulation platforms are difficult to fully replicate the details of physical interaction in the real world, causing strategies trained in simulation to experience sharp performance drops in real scenarios. Especially in contact-intensive tasks such as fine manipulation and flexible object interaction, the effect of sim-to-real transfer still has significant room for improvement. Therefore, improving long-horizon task generalization requires not only algorithm-level innovation but also the construction of a standardized dataset system, high-fidelity simulation environments, and a closed-loop iterative learning framework of “simulation pre-training–real fine-tuning–data回流.”
- Safety and Standardization
The safety assurance system of wheeled humanoid robots is key to their large-scale application. It currently faces multiple challenges such as physical interaction safety, data privacy protection, and lack of ethical norms. In terms of physical safety, collision risk prevention and control in human-robot coexistence scenarios is the core. Robots need intrinsic safety design and real-time obstacle avoidance capabilities. Although existing flexible drive and force control algorithms can reduce harm, in highly dynamic scenarios, the system still has deficiencies in collision identification, emergency stop, and safety strategy switching, and there is still a lack of unified, verifiable safety evaluation indicators. In terms of data privacy, sensitive information such as user images and voice collected by robots poses a risk of leakage, and dedicated encryption schemes for multimodal data are still immature. In terms of ethical safety, the allocation of decision-making authority in emergency tasks and the division of responsibility for operational errors lack clear guidelines and legal norms, and there is certain controversy.
The lack of standards and industry specifications is another key bottleneck restricting large-scale application. Insufficient standardization of core components and software interfaces, ununified mechanical arm joint interfaces, sensor transmission protocols, and API interfaces increase integration costs. The lack of performance testing standards means there is no unified testing and evaluation method for indicators such as operation accuracy and load capacity, making it difficult to objectively reflect product performance. International standards for industrial robots are still difficult to fully cover such systems that combine mobility and manipulation characteristics, and dedicated standards are still being improved. Domestic standards also mostly focus on industrial scenarios, while standards for commercial services and household scenarios are relatively lagging. In the future, it is urgent to establish a unified standard system covering multiple scenarios and the entire life cycle to support the standardized development and engineering implementation of wheeled humanoid robots.
- Core Components and System Integration
- Future Trends and Industry Perspectives
The engineering implementation of wheeled humanoid robots faces a core contradiction between improving general manipulation capability and ensuring scenario-specific reliability constraints. In response to this contradiction, frontier breakthroughs in the field are no longer limited to single-point iteration of hardware performance or algorithm effects. Instead, they form a full-chain evolution direction from underlying data and simulation systems, core intelligence and manipulation capabilities, new perception hardware, to safety and human-robot symbiosis paradigms, centered on the core characteristics of mobile-manipulation integration.
- Multimodal Data Collection, Processing, Training, and Simulation
In the field of embodied intelligence, processing massive data and achieving efficient learning has become a key technical bottleneck. The frontier trend is increasingly focusing on simulation-first, data-driven, and virtual-real closed loops. The core path is to use high-fidelity physical simulation platforms to build large-scale parallel virtual training grounds. Through domain randomization and domain adaptation technologies, visual and dynamic diversity as realistic as possible is injected into synthetic data, thereby pre-training strategies to learn robustness against the complexity and uncertainty of the real world. This process not only greatly reduces reliance on high-risk, high-cost trial and error in the physical world but also achieves order-of-magnitude improvements in training efficiency through distributed reinforcement learning architectures. However, relying solely on simulation cannot fully bridge the “gap” between simulation and reality. Therefore, frontier research is committed to building automated, closed-loop data pipelines: deploying strategies pre-trained in simulation onto real robots for “silent trials” or human-supervised interaction. The collected real-world trajectories are automatically cleaned and annotated to form high-value datasets. These data can be directly used for online imitation learning or offline reinforcement learning to fine-tune strategies. On the other hand, through a simulation-to-real-to-simulation loop, they can be used to correct the physical parameters of the simulation model, making the virtual environment continuously approach the real world.
The ultimate form of this trend is to build a robot digital twin system that is synchronized with the physical entity in real time and has predictive capability. It makes it possible to safely and rapidly test new skills, plan complex task chains, and rehearse anomaly recovery plans in virtual space, and then seamlessly issue verified plans to the physical entity for execution. By boldly exploring in simulation and executing robustly in reality, the development, testing, and deployment cycle of robot skills is greatly compressed, enabling robot learning to shift comprehensively from the traditional model relying on manual programming and a small number of experiments to a scalable new paradigm driven by both large-scale synthetic data and continuous real feedback.
- General Intelligence and Human-Like Dexterous Manipulation
General intelligence and human-like dexterous manipulation are core development directions in the field of wheeled humanoid robots, accelerating from traditional specialized skill-driven approaches to general embodied intelligence empowered by large models. Early robot dexterous manipulation relied on task-specific models with limited generalization ability. General embodied intelligence based on large language models integrates multimodal perception information such as vision, touch, and force perception to achieve end-to-end mapping from natural language instructions to manipulation tasks, improving the naturalness of human-robot interaction and the flexibility of task adaptation. Combined with collaborative training paradigms of reinforcement learning and sub-skill controllers, robots can autonomously acquire human-like manipulation behaviors such as finger gaiting and multi-finger coordination. The core is to optimize exploration efficiency in complex state spaces through value-guided exploration strategies, allowing high-precision dexterous manipulation capabilities to emerge without manual demonstration. At the hardware level, structural optimization and sensor integration of multi-DOF dexterous hands provide the physical foundation. Domain randomization technology in simulation environments narrows the sim-to-real gap, enabling trained strategies to be stably deployed on physical robots. This collaborative evolution path drives dexterous manipulation from task-specific to environment-adaptive, instruction-self-understanding general intelligence, laying a key foundation for large-scale application of wheeled humanoid robots in complex scenarios.
- Novel Sensors Through Multidisciplinary Integration
With the development of materials science and neural engineering, multidisciplinary integration of novel sensors is giving robots more advanced perception capabilities. In addition to traditional force perception and vision, “electronic skin” based on flexible electronics technology can give robots large-area tactile perception, enabling safe collision detection and fine manipulation capabilities in human-robot interaction. In addition, the introduction of electromyography (EMG) interfaces and brain-machine interfaces (BMI) allows robots to directly parse human motion intentions. This integration of intention recognition based on biological signals and novel sensors will greatly expand the application boundaries of dual-arm humanoid robots in medical rehabilitation and human-robot collaboration.
Furthermore, under the trend of synchronous evolution in perception and expression, facial simulated expressions are gradually becoming an important extension of multimodal sensing and interaction. The development of flexible, stretchable sensing and actuation materials allows robots to integrate high-density “facial electronic skin” on their faces, which can both sense touch, pressure, and temperature and present delicate facial expression changes through deformable structures. These facial expression signals, combined with vision, voice, touch, and bioelectrical signals, can be incorporated into a multimodal perception-decision framework for emotion recognition and emotional expression, thereby enhancing the robot’s social presence and long-term coexistence acceptability in scenarios such as companionship, education, and rehabilitation.
Through the deep integration of multi-source novel sensors such as electronic skin, bioelectrical interfaces, and facial simulated expressions, wheeled humanoid robots are gradually moving from “physical collaboration tools” to “intelligent partners” with situational understanding and emotional interaction capabilities. This will greatly expand their application boundaries in medical rehabilitation, human-robot collaborative manufacturing, and daily services.
- Safety Ethics and Human-Robot Symbiosis Paradigm
Human-robot-environment symbiosis and safety ethics standardization are core supports for wheeled humanoid robots to practice the concepts of “human-centered” and “human-centric intelligent manufacturing” and to build a new paradigm of human-robot symbiosis. As their application scenarios extend to all sectors of society, achieving harmonious coexistence among humans, robots, and the environment has become a key proposition for technological iteration and industrial development. To achieve this goal, the safety design and ethical norm construction of wheeled humanoid robots are being comprehensively advanced from three dimensions: technology, institutions, and society, gradually forming a development system that fits their mobile-manipulation collaborative characteristics.
At the technical level, safety optimization focuses on risk prevention and control in both mobile and manipulation scenarios. The deep combination of multimodal sensor fusion and intelligent decision-making algorithms enables robots to accurately perceive potential risks in dynamic environments and avoid safety hazards such as collisions and overloads by adjusting mobile trajectories and manipulation forces in real time. Intrinsic safety design for mobile-manipulation coupling continues to upgrade. The application of flexible drive technology and adaptive control strategies gives robots better compliance and fault tolerance in human-robot interaction, reducing physical interaction risks from the underlying architecture and laying a solid technical foundation for the implementation of human-centric intelligent manufacturing.
At the ethical level, as wheeled humanoid robots participate in more complex tasks, issues such as decision-making boundaries, responsibility division, and data privacy protection have attracted widespread attention. Governments, international organizations, and academia are jointly promoting the formulation of targeted norms, establishing ethical guidelines and legal frameworks adapted to their technical characteristics around core issues such as behavioral constraints in mobile operations, data collection and use specifications, and fault responsibility definition, ensuring that robot behavior always conforms to social ethics and public interest.
At the social level, deepening human-robot symbiosis depends not only on technical and institutional guarantees but also on the cognitive upgrade of society as a whole regarding wheeled humanoid robots. Through science popularization and application practice, a public trust system for robots is gradually established. At the same time, the positioning of robots in the social division of labor is clarified to ensure the controllability and transparency of their behavior, achieving synchronous improvement of technological progress and social acceptance. With the continuous improvement of safety and ethical systems, wheeled humanoid robots will integrate more smoothly into human production and life, deeply practicing the concepts of “human-centered” and “human-centric intelligent manufacturing.” Driven by both technological innovation and social consensus, the new paradigm of human-robot symbiosis will continue to mature, injecting lasting momentum into industrial upgrading and sustainable social development.
- Multimodal Data Collection, Processing, Training, and Simulation
- Industry Development Considerations and Recommendations
As a core engineering carrier for the implementation of embodied intelligence technology, the industrial development of wheeled humanoid robots is deeply coupled with the evolution of the technological frontier. It is currently in a critical window period transitioning from technical validation to application deployment. A preliminary industrial prototype covering core components, system integration, and application validation has formed, but it still faces practical problems such as poor transformation of core technologies, insufficient depth of scenario deployment, and an incomplete industrial support system. It is necessary to build an industrial development path adapted to technical characteristics and market needs around three core dimensions: technological innovation, scenario deployment, and ecosystem construction.
First, strengthen collaborative technological innovation to break through core industrial bottlenecks. Focusing on the strongly coupled technical characteristics of mobile manipulation in wheeled humanoid robots, efforts should focus on breaking through core algorithms such as unified dynamics modeling, whole-body motion planning for high-redundancy systems, and collaborative control in dynamic environments. At the same time, key hardware such as high-power-density miniaturized servo drives, high-density flexible tactile sensing, and highly reliable dexterous hand integration should be tackled. Overall energy management and lightweight design should be optimized to break the core contradiction among performance, volume, and energy consumption. Relying on a collaborative innovation system with deep industry-university-research integration, integrate the basic research accumulation of universities and research institutions, the engineering and scenario validation advantages of leading enterprises, build special innovation platforms, and open up the transformation chain of technological achievements from laboratory to industrialization. Build public infrastructure such as open-source simulation training platforms and standardized embodied manipulation datasets, unify data collection and annotation specifications, reduce industry innovation thresholds, and promote the industry to achieve a leap from single-point technological breakthroughs to system-level capability improvement.
Second, promote deep scenario penetration and innovate application service models. The technical value and industrial value of wheeled humanoid robots ultimately need to be realized through scenario deployment. It is necessary to follow the principles of gradient penetration and demand orientation to build a ladder-style scenario expansion path. It is recommended to prioritize deep cultivation of core industrial manufacturing scenarios, targeting industries and scenarios such as new energy vehicles, 3C electronics, logistics sorting, patrol and guard, and inspection. Focus on clear needs such as collaborative assembly, material handling, and equipment inspection to create standardized solutions adapted to production line cycles and consolidate the basic plate of industrial profitability. At the same time, steadily expand commercial service scenarios such as retail, hotels, medical rehabilitation, and elderly and disability assistance, optimize human-robot interaction safety and multi-task adaptability, and create replicable benchmark demonstration projects. Prospectively layout household service scenarios and promote high-reliability product iteration. Adopt a product development model of “general-purpose onboard platform plus modular configuration plus scenario-based process package” to balance generalization and customization needs. Simultaneously promote new business models such as RaaS and leasing plus operation and maintenance integration, building a commercial path that links hardware, software, services, and data.
Third, improve the industrial ecosystem and build a solid foundation for industrial development. A sound industrial ecosystem is the core support for wheeled humanoid robots to achieve large-scale, high-quality development. It is necessary to build a collaborative and symbiotic industrial development pattern around three core directions: standard system construction, industrial chain collaboration, and talent supply. Accelerate the construction of an industry standard system covering the entire product life cycle, unify hardware interfaces, communication protocols, and software API specifications, improve product performance testing methods and evaluation systems, formulate special standards for physical safety and data privacy safety in human-robot coexistence scenarios, and actively participate in international standard formulation to enhance international discourse power in this field. Deepen collaboration across the entire industrial chain, cultivate an industrial pattern led by leading enterprises and supported by specialized and sophisticated small and medium-sized enterprises, strengthen the domestic supply capacity of core components, ensure the autonomy and controllability of the supply chain, and improve the overall resilience of the industrial chain. Build a multi-level composite talent training system, and through interdisciplinary construction, university-enterprise joint training, and vocational skills training, create a talent echelon covering the entire chain of research and development, engineering, and application, providing stable intellectual support for the sustainable development of the industry.
- Conclusion
Research on wheeled humanoid robots has formed a relatively complete technical and industrial development context and framework. The core covers onboard hardware structure and software architecture, mobile chassis environment perception and motion planning, task understanding and dual-arm dexterous manipulation, data collection and simulation training for complex scenarios, and embodied intelligence control paradigms and sim-to-real transfer. The current research focus is on unified dynamics modeling of the wheeled chassis and dual-arm system, collaborative motion planning for high-redundancy systems, multimodal sensor information fusion, and compliant control. The technological evolution shows a clear trend from independent control of subsystems to integrated whole-body control of mobility and manipulation. The integrated application of multi-sensor fusion, high-fidelity simulation, and intelligent algorithms has become a core path to breaking existing technical bottlenecks and also provides important equipment support for the physical implementation of embodied intelligence.
Wheeled humanoid robots have completed engineering pilot applications in industrial manufacturing, commercial services, special operations, and other fields. A development pattern of parallel advancement in technology research and engineering implementation has formed, and a preliminary industrial system covering core components, complete machine research and development, and scenario applications has been established. The current large-scale engineering application is still limited by high-performance core component integration, collaborative planning in dynamic environments, dataset collection and training, task generalization in unstructured scenarios, and an incomplete industry standard system. In the future, with continuous breakthroughs in related technologies and the continuous improvement of the industrial ecosystem, wheeled humanoid robots will achieve broader large-scale applications in intelligent manufacturing and service fields, possessing important academic research value and engineering application prospects. The continued advancement of embodied intelligence will be central to this transformation, enabling wheeled humanoid robots to become reliable, adaptive, and collaborative partners in an increasingly automated and intelligent world.