A newly reported investigation into embodied intelligence has examined how expert strategy organization and controlled environmental conditions during demonstration data collection affect model performance in complex robotic manipulation. The work introduces a dual ViA segmented sampling strategy, supported by controllable illumination, background, and tabletop settings, and evaluates the resulting datasets on a dual-arm collaborative robot platform.
The task used for validation was an industrial raw-material grasping and placing operation with clutter interference. Performance was measured through grasp success rate, placement success rate, and task execution position deviation. According to the reported results, combining dual ViA segmented sampling with controlled illumination, background, and tabletop conditions increased grasp success rate by 52%, placement success rate by 61%, and reduced task execution position deviation by 70% compared with data collected using dual ViA segmented sampling under conventional environmental conditions.
The findings place embodied intelligence data quality at the center of robotic learning. Rather than treating demonstration collection as a neutral recording process, the study organizes expert behavior into explicit task stages and simultaneously constrains environmental variables that shape visual input. The result is a structured dataset that preserves continuous task execution while making critical manipulation phases more visible to the learning model. This approach addresses a persistent challenge in embodied intelligence: the gap between large-scale standardized datasets and the long-sequence, high-precision, task-specific data required for industrial operation.
1. Embodied Intelligence Shifts from Standardized Datasets toward Task-Specific Demonstration Quality
The rapid development of embodied intelligence has increased expectations for robots operating in flexible manufacturing, autonomous task planning, and complex environments. Robotic systems are moving from program-controlled automation toward embodied agents capable of perception, cognition, and decision-making. However, the performance of embodied intelligence models depends not only on model architecture and training algorithms but also on the quality of demonstration datasets. These datasets must provide multimodal information, broad task coverage, sampling precision, and data consistency for both training and validation.
Existing embodied intelligence datasets have largely concentrated on household kitchens, daily life, industrial production, and healthcare scenarios. Representative efforts include Open X-Embodiment, which spans 22 embodiments and 160,000 task demonstrations under a unified robot learning dataset standard; RoboMIND, which contains 107,000 demonstration trajectories across 479 tasks and 96 object categories, including 5,000 failure cases; ARIO, which integrates real and simulated data from 258 robot platforms and provides 3 million demonstrations and 321,064 tasks; and Kaiwu, which offers 11,664 multimodal demonstration instances for complex assembly tasks. These resources have advanced cross-embodiment learning and multimodal perception.
At the same time, the study notes that many existing datasets focus on standardized scenes and simplified tasks. They may not directly satisfy the demands of specific industrial operations, especially long-sequence, high-precision dual-arm manipulation. In embodied intelligence, visual imitation learning has shown that policy generalization can be affected by changes in illumination, camera position, and other environmental conditions. In addition, existing data collection methods often lack explicit organization of key stages in long-sequence tasks, allowing critical operation samples to be diluted by transition segments. Changes in illumination, background, and tabletop conditions can shift the visual input distribution, affecting target recognition, action alignment, and placement localization. These factors motivate a task-specific collection framework that combines expert strategy organization with controllable environmental settings.
2. A Dual ViA Segmented Sampling Strategy for Structured Demonstration Collection
The proposed method introduces a dual ViA segmented sampling strategy for dual-arm robot manipulation. ViA refers to Vision in Action, an approach that emphasizes active perception and synchronized observation-action behavior. In this study, dual ViA means that both the observation side and the execution side are coordinated within a staged collection process. The strategy does not change the continuous execution flow of the robot. Instead, it reorganizes the data collection process so that the task is divided into semantically meaningful stages, and the recorded observations, states, gripper states, master-side control inputs, executed actions, and stage labels are synchronized under a unified timestamp.
The grasping and placing task is divided into eight stages: target localization, approach alignment, gripper closure, lift and transfer, descent alignment, gripper release, lift avoidance, and arm reset. The robot still performs the task continuously, but the data recording process is organized according to stage labels. This makes critical operations such as grasp alignment, stable clamping, and placement alignment explicitly identifiable within a long sequence. As a result, the influence of ineffective transition fragments on policy learning is reduced.
On the observation side, the system uses the position deviation of the target in the image plane as feedback to drive small adjustments of a wrist-mounted camera. The goal is to keep the target near the center of the image. Smoothing constraints suppress abrupt viewpoint changes and reduce image jitter, preserving temporal observation continuity. The system also constrains camera observation distance and end-effector pose so that the target and critical operation area remain within the effective field of view. Observation-side control is mainly applied during target localization, lift and transfer, descent alignment, gripper release, and arm reset, improving the stability of visual input in key stages.
On the execution side, the system organizes key actions in a stepwise manner. During grasping, the robot performs approach alignment and then gripper closure, followed by a brief pause after the gripper closes to stabilize clamping depth and contact state. During placing, the robot performs descent alignment, gripper release, lift avoidance, and arm reset in sequence, clarifying the action objective of each stage. The execution process prioritizes vertical descent, vertical lift, and smooth transitions to reduce the effects of speed突变, pose deviation, and contact disturbance on demonstration trajectories.
The method defines a task stage state set and uses a stage determination function based on target visibility, relative position between the end effector and the target, gripper closure state, and task completion status. Observation-side and execution-side control triggers are determined by the current stage. The observation side is triggered in target localization, lift and transfer, descent alignment, gripper release, and arm reset. The execution side is triggered in approach alignment, gripper closure, lift and transfer, descent alignment, gripper release, lift avoidance, and arm reset. Control outputs are generated by separate functions for observation and execution.
For data organization, the algorithm synchronizes observations, robot states, gripper states, master-side control inputs, actions, and stage labels under a common timestamp. A structured sampling unit is represented by the observation, state, gripper state, master control input, executed action, and stage label. A complete episode consists of a sequence of these units. By introducing structured sampling units and stage labels, the system preserves the complete task flow while explicitly marking the positions of key stages in the sequence. This allows subsequent models to identify critical operation processes more clearly and reduces interference from ineffective transition segments.
The dual ViA segmented scheduling algorithm is not intended to alter the task execution order. Its role is to establish a unified stage-driven scheduling mechanism for long-sequence demonstration collection. It coordinates observation-side viewpoint control and execution-side action constraints around stage boundaries, and it completes structured organization of multimodal data under continuous execution. This turns the dual ViA segmented sampling strategy into an executable and synchronizable collection process, improving the validity, consistency, and reproducibility of demonstration data for embodied intelligence.
3. Environmental Condition Control for Illumination, Background, and Tabletop
The study treats environmental conditions as controllable, quantifiable, and reproducible variables rather than random contextual noise. In embodied intelligence, visual input is highly sensitive to illumination, background, and surface conditions. If these factors fluctuate across collection batches, the resulting dataset may contain distribution shifts that degrade training and validation comparability. The proposed method therefore controls three environmental dimensions: illumination, background, and tabletop.
3.1 Illumination Condition Control
Illumination directly affects camera imaging quality and depth observation stability. Fluctuations over time, weather, or external surroundings can introduce shadows, reflections, and uneven brightness. To reduce these effects, the system places shadowless lamps and top LED array lights above the work area and uses reflective cloth to unify the illumination range and brightness distribution. The shadowless lamps reduce local shadows and high-reflection interference in the target area, while the LED array lights improve overall illumination uniformity.
To quantitatively assess illumination uniformity, the study uses the coefficient of variation as a measure of illumination intensity fluctuation. The coefficient of variation is calculated as the ratio of the standard deviation of illumination values in a sampling region to the mean illumination value. A smaller coefficient indicates more uniform illumination and more stable visual observation. The controlled illumination setting is intended to reduce this coefficient, suppress shadow and reflection interference, and improve temporal observation stability during grasping and placing.
3.2 Background Condition Control
Background conditions affect the distinguishability of target objects. If background color, texture, or complexity changes between collection batches, target boundaries may become blurred, regional contrast may decrease, and visual input distribution may shift. The proposed background control installs enclosures on three sides of the work area and covers them with background cloth of uniform color and material. This constrains background color and texture features and reduces visual complexity.
To measure the similarity between target and background regions, the study uses histogram matching. The matching value is calculated from the histograms of the background image and the target image across color components. A larger matching value indicates a higher proportion of the same color between background and target, meaning the target and background are more similar and the target boundary is less distinguishable. A smaller matching value indicates clearer color difference and stronger target distinguishability. The controlled background setting aims to reduce the matching value, suppress background interference with target boundary recognition, and improve target recognition stability in embodied intelligence models.
3.3 Tabletop Condition Control
Tabletop conditions affect target distinguishability and visual observation stability. Changes in tabletop material, color, or texture can cause local reflection, boundary noise, and brightness contrast fluctuations. These factors may blur target boundaries, shift visual input distribution, and create inconsistent spatial references during placement. The proposed tabletop control uses a uniform covering layer with consistent material and color. This reduces local highlight regions and surface noise caused by metal surfaces.
To quantify the brightness contrast between target and tabletop, the study uses Weber contrast. The contrast value is calculated from the average brightness of the target object region and the average brightness of the tabletop region. When the tabletop brightness is approximately constant, the formula characterizes the brightness contrast of the target relative to the tabletop. A larger absolute contrast value indicates a clearer brightness difference and a clearer target boundary. The controlled tabletop setting aims to increase the absolute Weber contrast, making the target boundary clearer and improving visual observation stability.
An algorithm for tabletop condition control adjusts tabletop material, texture complexity, color, or brightness if reflection, texture complexity, or contrast does not satisfy predefined thresholds. It repeats acquisition and evaluation until reflection is below the threshold, texture complexity is below the threshold, and absolute Weber contrast is above the threshold. This procedure ensures that the tabletop configuration meets the collection requirements before demonstration data are recorded.
4. Experimental Design on a Dual-Arm Collaborative Robot Platform
The experiments were conducted on a Realman dual-arm collaborative robot platform. The platform was equipped with 4 cameras as visual input devices: left and right wrist cameras, a top camera, and a bottom camera. All experiments used the ACT model, or Action Chunking with Transformers, as the unified baseline model. ACT predicts multi-step action chunks and is suitable for long-sequence control tasks such as dual-arm manipulation. The study used ACT to compare how datasets collected under different demonstration strategies and environmental conditions affect model validation performance.
The task was industrial raw-material grasping and placing with clutter interference. A precision requirement of 2.5 mm was set. The industrial raw material was allowed to shift and rotate within a specified range, while the vise maintained fixed size, position, and direction. Two placement conditions were tested. In condition 1, the initial position and direction of the industrial raw material were randomly placed within a center radius of 5 cm. In condition 2, the initial position and direction were randomly placed within a center radius of 10 cm. The experiments conducted 50 online inferences in each condition. After averaging the results of condition 1 and condition 2, the study recorded grasp success rate, placement success rate, and task execution accuracy. Task execution accuracy was represented by target placement position deviation, where a smaller value indicates higher accuracy.
Clutter was introduced to simulate complexity in real environments. The clutter consisted of 3 red building blocks and 2 yellow building blocks. The red blocks measured 40 mm × 20 mm × 50 mm, and the yellow blocks measured 40 mm × 20 mm × 40 mm. Clutter positions were generated through a stratified random sampling mechanism. The procedure randomly selected the raw material position and then generated clutter positions under a minimum distance constraint from the raw material. This ensured random distribution while avoiding contact between the gripper and clutter during grasping, so the clutter mainly formed visual interference.
Six datasets were used for embodied intelligence model training and inference validation. Each dataset contained 100 demonstrations. The datasets were designed as follows:
- Conventional illumination, conventional background, and conventional tabletop conditions with traditional data collection, serving as the baseline dataset.
- Conventional illumination, conventional background, and conventional tabletop conditions with dual ViA segmented sampling, used to analyze the collection strategy.
- Controlled illumination, conventional background, and conventional tabletop conditions with dual ViA segmented sampling, used to analyze illumination effects.
- Conventional illumination, controlled background, and conventional tabletop conditions with dual ViA segmented sampling, used to analyze background effects.
- Conventional illumination, conventional background, and controlled tabletop conditions with dual ViA segmented sampling, used to analyze tabletop effects.
- Controlled illumination, controlled background, and controlled tabletop conditions with dual ViA segmented sampling, used to analyze joint environmental control.
The conventional illumination condition relied only on indoor lighting and natural light, simulating common natural illumination without additional artificial light sources. The conventional background condition had no surrounding遮挡, and the worktable was rotated after a certain amount of data collection to ensure background inconsistency. The conventional tabletop condition used a metal work surface, simulating common material and color in practical environments.
In the dual ViA segmented sampling operation flow, the right arm’s ViA was mainly responsible for visual tracking and localization of the target raw material, dynamically adjusting the viewpoint to optimize visual input quality. The left arm’s ViA was responsible for grasping and placing, adjusting actions according to target pose changes to ensure precise grasping and placement. The segmented sampling strategy divided the long-sequence task into small intervals and used precise timestamp synchronization to align data across stages. If data collection exceeded a predetermined step length, the data were considered invalid and not used for model training.
The staged process included target localization and view alignment, approach and grasp preparation, gripper closure, lift and transfer, descent into the vise groove, gripper release, lift avoidance, and return to initial position. The traditional data collection process kept the right arm static and used only the left arm to perform grasping and placing. It did not explicitly divide the long sequence into key stages and lacked stable multi-view observation. As a result, observation and action information in key operations were mixed, reducing the representational capacity of the collected data for critical stages.
5. Results: Expert Strategy and Environmental Control Improve Embodied Intelligence Performance
The study designed five groups of comparative experiments to evaluate expert strategy data collection and different environmental conditions. The experiments included a comparison between traditional data collection and dual ViA segmented sampling, three single-factor comparisons for illumination control, background control, and tabletop control, and a comparison between joint environmental control and no control. The analysis used grasp success rate, placement success rate, and task execution position deviation as indicators of model inference performance under different collection conditions.
5.1 Expert Strategy Data Collection Comparison
In the expert strategy comparison, illumination, background, and tabletop conditions remained conventional. Only the data collection method was changed. The dual ViA segmented sampling strategy achieved an average grasp success rate of 34%, an average placement success rate of 18%, and an average accuracy of 2.5 mm. Compared with the traditional data collection method, grasp success rate increased by 22 percentage points and placement success rate increased by 13 percentage points. Task execution position deviation remained unchanged.
The condition-level results are shown in the following data table.
| Method | Condition 1 Grasp Success (%) | Condition 1 Placement Success (%) | Condition 1 Accuracy (mm) | Condition 2 Grasp Success (%) | Condition 2 Placement Success (%) | Condition 2 Accuracy (mm) |
|---|---|---|---|---|---|---|
| Traditional method | 22 | 10 | 2.5 | 2 | 0 | 2.5 |
| Dual ViA method | 46 | 28 | 2.5 | 22 | 8 | 2.5 |
Statistical significance was assessed using a two-proportion Z-test. For grasp success rate, the p-value was 0.0113 in condition 1 and 0.0021 in condition 2. For placement success rate, the p-value was 0.0218 in condition 1 and 0.0412 in condition 2. All p-values were below 0.05, indicating that the performance improvements associated with dual ViA segmented sampling were statistically significant.
The results indicate that traditional data collection had limitations for long-sequence dual-arm manipulation. The right arm remained static, and the left arm independently completed grasping and placing. The collection process did not divide the task into key stages, and stable multi-view observation was absent. Consequently, observation and action information in critical operations were mixed, and the collected data had insufficient representational capacity for key stages. In contrast, dual ViA segmented sampling organized the grasping and placing process according to task semantics and operation objectives. Observation-side viewpoint control, execution-side action constraints, and stage-driven scheduling kept observation, state, and action information in stable correspondence during critical stages. This expanded effective visual coverage during task execution, strengthened the representation of key-stage samples, improved demonstration data consistency and validity, and produced better task execution stability during model inference.
5.2 Illumination Condition Comparison
In the illumination comparison, background and tabletop conditions remained conventional, and the dual ViA segmented sampling method was used. Only illumination was controlled. Under controlled illumination, the model trained on dual ViA data achieved an average grasp success rate of 65%, an average placement success rate of 46%, and an average accuracy of 1.75 mm. Compared with the conventional illumination condition, grasp success rate increased by 31 percentage points, placement success rate increased by 28 percentage points, and task execution position deviation decreased by 30%.
| Illumination Condition | Condition 1 Grasp Success (%) | Condition 1 Placement Success (%) | Condition 1 Accuracy (mm) | Condition 2 Grasp Success (%) | Condition 2 Placement Success (%) | Condition 2 Accuracy (mm) |
|---|---|---|---|---|---|---|
| Conventional illumination | 46 | 28 | 2.5 | 22 | 8 | 2.5 |
| Controlled illumination | 78 | 52 | 1.5 | 52 | 40 | 2.0 |
The two-proportion Z-test showed statistically significant improvements. For grasp success rate, the p-value was 0.00098 in condition 1 and 0.0019 in condition 2. For placement success rate, the p-value was 0.0143 in condition 1 and 0.00018 in condition 2. All p-values were below 0.05.
The study explains that under conventional conditions, work-area illumination is affected by natural light changes, and the robotic arms can produce occlusion shadows during grasping and placing. This creates uneven illumination distribution and increases the coefficient of variation, reducing visual input stability. Under controlled illumination, the unified configuration of white reflective cloth, shadowless lamps, and top LED array lights makes work-area illumination more uniform. The coefficient of variation decreases明显, indicating that illumination fluctuation and local shadow interference are effectively suppressed. The imaging quality of the target area and the stability of temporal observation improve, showing that illumination uniformity is an important factor affecting the execution stability of grasping and placing tasks in embodied intelligence.
5.3 Background Condition Comparison
In the background comparison, illumination and tabletop conditions remained conventional, and dual ViA segmented sampling was used. Only background conditions were controlled. Under controlled background, the model achieved an average grasp success rate of 59%, an average placement success rate of 39%, and an average accuracy of 1.75 mm. Compared with conventional background, grasp success rate increased by 25 percentage points, placement success rate increased by 21 percentage points, and task execution position deviation decreased by 30%.
| Background Condition | Condition 1 Grasp Success (%) | Condition 1 Placement Success (%) | Condition 1 Accuracy (mm) | Condition 2 Grasp Success (%) | Condition 2 Placement Success (%) | Condition 2 Accuracy (mm) |
|---|---|---|---|---|---|---|
| Conventional background | 46 | 28 | 2.5 | 22 | 8 | 2.5 |
| Controlled background | 72 | 46 | 1.5 | 46 | 32 | 2.0 |
Statistical analysis showed that grasp success rate improvements were significant in both conditions, with p-values of 0.0082 in condition 1 and 0.0113 in condition 2. Placement success rate was significant in condition 2 with a p-value of 0.0027, but not in condition 1 with a p-value of 0.0623.
Under conventional background conditions, background color and texture lack unified constraints, and the color distribution similarity between target and background is relatively high. The histogram matching value is relatively large, making target boundaries less prominent and affecting the model’s recognition of target position. Under controlled background conditions, unified background color and texture reduce color similarity between target and background. The histogram matching value decreases明显, indicating that background interference with target boundary recognition is effectively suppressed. The contour and position features of the target object become more prominent, suggesting that background consistency helps reduce visual interference and improve target recognition stability in embodied intelligence.
5.4 Tabletop Condition Comparison
In the tabletop comparison, illumination and background conditions remained conventional, and dual ViA segmented sampling was used. Only tabletop conditions were controlled. Under controlled tabletop, the model achieved an average grasp success rate of 51%, an average placement success rate of 27%, and an average accuracy of 2.25 mm. Compared with conventional tabletop, grasp success rate increased by 17 percentage points, placement success rate increased by 9 percentage points, and task execution position deviation decreased by 10%.
| Tabletop Condition | Condition 1 Grasp Success (%) | Condition 1 Placement Success (%) | Condition 1 Accuracy (mm) | Condition 2 Grasp Success (%) | Condition 2 Placement Success (%) | Condition 2 Accuracy (mm) |
|---|---|---|---|---|---|---|
| Conventional tabletop | 46 | 28 | 2.5 | 22 | 8 | 2.5 |
| Controlled tabletop | 60 | 34 | 2.0 | 42 | 20 | 2.5 |
For grasp success rate, the p-value was 0.161 in condition 1 and 0.0321 in condition 2. For placement success rate, the p-value was 0.517 in condition 1 and 0.0838 in condition 2. Grasp success rate was statistically significant in condition 2 but not in condition 1. Placement success rate was not statistically significant in either condition.
Under conventional tabletop conditions, the metal surface easily produces local reflection and boundary noise. The brightness contrast between the target object and the tabletop region is insufficient, and the absolute Weber contrast is relatively small, reducing target boundary clarity. Under controlled tabletop conditions, unified tabletop material and surface color enhance brightness contrast between target and tabletop. The absolute Weber contrast increases明显, indicating that local reflection and boundary noise are effectively suppressed. The target boundary becomes clearer during the placement stage, showing that tabletop condition control mainly improves positioning stability by improving the target-tabletop visual contrast relationship. For embodied intelligence, this result suggests that tabletop control is meaningful but may have a smaller isolated effect than illumination or background control in the tested task.
5.5 Joint Environmental Control Comparison
In the joint environmental control experiment, the study compared conventional conditions with controlled illumination, controlled background, and controlled tabletop conditions while using dual ViA segmented sampling. Under joint control, the model achieved an average grasp success rate of 86%, an average placement success rate of 79%, and an average accuracy of 0.75 mm. Compared with conventional conditions, grasp success rate increased by 52 percentage points, placement success rate increased by 61 percentage points, and task execution position deviation decreased by 70%.
| Condition | Condition 1 Grasp Success (%) | Condition 1 Placement Success (%) | Condition 1 Accuracy (mm) | Condition 2 Grasp Success (%) | Condition 2 Placement Success (%) | Condition 2 Accuracy (mm) |
|---|---|---|---|---|---|---|
| Conventional conditions | 46 | 28 | 2.5 | 22 | 8 | 2.5 |
| Joint controlled conditions | 90 | 84 | 0.5 | 82 | 74 | 1.0 |
The two-proportion Z-test showed strong statistical significance. For grasp success rate, the p-value was 2.40×10-6 in condition 1 and 1.92×10-9 in condition 2. For placement success rate, the p-value was 1.69×10-8 in condition 1 and 1.95×10-11 in condition 2. All p-values were far below 0.05.
Under conventional conditions, illumination intensity is uneven, and the robotic arms produce shadows during grasping and placing, causing illumination fluctuation that interferes with model execution. The visual difference between background and object is small, reducing object visual salience and lowering recognition accuracy. The visual difference between tabletop and object is also small, making object edges unclear and affecting recognition accuracy and positioning precision. These factors together lead to unstable task execution. When illumination, background, and tabletop conditions are controlled simultaneously, the originally random environmental conditions are transformed into controllable, quantifiable, and reproducible variables. Work-area illumination distribution becomes more stable, and visual distinction between target and background, as well as between target and tabletop, increases simultaneously. The interference of the three environmental conditions on visual input is suppressed as a whole. Compared with single-factor environmental control, the model achieves better grasp and placement success rates, indicating that joint environmental control can more effectively improve the consistency and stability of visual input during demonstration data collection. This further improves inference stability in the raw-material grasping and placing task for embodied intelligence.
6. Why the Gains Matter for Embodied Intelligence
The reported results have several implications for embodied intelligence research and industrial deployment. First, they show that expert strategy organization is not merely a matter of collecting more data. The explicit segmentation of long-sequence tasks changes the internal structure of the dataset. By labeling target localization, approach alignment, gripper closure, lift and transfer, descent alignment, gripper release, lift avoidance, and arm reset, the dataset preserves the complete task while making critical stages easier to learn. This is especially important for dual-arm manipulation, where coordination, timing, and contact stability determine success.
Second, the results show that environmental control can improve embodied intelligence model validation performance without changing the underlying model architecture. Illumination, background, and tabletop conditions are often treated as background variables in data collection. The study demonstrates that these variables can be measured and regulated using the coefficient of variation, histogram matching, and Weber contrast. When these variables are controlled, visual input distribution becomes more stable across collection batches. For embodied intelligence models, this stability supports better target recognition, action alignment, and placement localization.
Third, the joint control experiment suggests that environmental factors interact. Illumination control alone, background control alone, and tabletop control alone each produced improvements in some metrics. However, the combined control produced the strongest gains. This indicates that embodied intelligence systems operating in complex scenes may benefit from a holistic visual environment design rather than isolated adjustments. In industrial settings, such a design could include standardized lighting, uniform background enclosures, and consistent tabletop surfaces as part of the data collection infrastructure.
Fourth, the study reinforces the importance of task-specific data for embodied intelligence. Large standardized datasets remain valuable for generalist robot learning. However, industrial tasks often involve long sequences, high precision, specific object properties, and clutter. The dual ViA segmented sampling strategy provides a way to organize expert demonstrations so that these task-specific requirements are represented in the data. The use of ACT as a baseline model also suggests that the benefits can be observed with a widely used action-chunking architecture, making the approach relevant to current embodied intelligence practice.
7. Limitations and Future Work
The study acknowledges that the proposed method still has room for optimization. Future work will focus on adaptive modeling and compensation for environmental uncertainty, including illumination changes, background disturbances, and tabletop material variations. The goal is to improve system stability under unstructured collection conditions. In addition, the method should be validated in more complex long-sequence tasks such as assembly and insertion. These tasks may introduce additional challenges, including contact-rich manipulation, tighter tolerances, and more varied object geometries.
For embodied intelligence, the next step is to move from controlled collection environments to adaptive collection systems that can maintain data quality when conditions change. This may require online monitoring of visual metrics, automatic adjustment of illumination and background settings, and dynamic selection of demonstration segments. The current study provides a structured baseline by showing how expert strategy organization and environmental control can improve model performance in a dual-arm industrial task.
8. Conclusion
The study addresses demonstration data collection for dual-arm collaborative robot grasping and placing. It proposes expert strategy data collection and environmental condition control methods for embodied intelligence. The dual ViA segmented sampling strategy explicitly identifies key task stages and synchronizes observation-side viewpoint control with execution-side action constraints under a unified timestamp, improving the consistency and validity of demonstration data. Illumination, background, and tabletop conditions are controllably set, and comparative datasets are constructed to analyze environmental effects on model validation performance.
Experimental results show that, compared with data collected using dual ViA segmented sampling under conventional environmental conditions, joint use of dual ViA segmented sampling and environmental condition control increased grasp success rate by 52%, increased placement success rate by 61%, and reduced task execution position deviation by 70%. These results indicate that expert strategy organization and environmental condition control can strengthen the representation of key-stage samples, reduce environmental interference with visual input, and improve the stability of model inference in complex scenes.
The findings support a broader direction for embodied intelligence: high-quality demonstration data depends not only on the amount of data collected but also on how expert behavior is structured and how environmental conditions are managed. For industrial robotics, where precision, reliability, and adaptation to complex scenes are critical, the combination of staged expert strategy collection and controllable visual environments offers a practical path toward more robust embodied intelligence models.