Humanoid robot technology has long captured the imagination of researchers and the public alike. The pursuit of creating machines that resemble and interact with humans is not merely an exercise in mimicry but a driving force behind advances in computer science, control theory, mechanics, information sensing, artificial intelligence, and bionics. Among the many subsystems that constitute a humanoid robot, the head plays a uniquely important role. It houses vision, hearing, and other sensory organs, and its motion capabilities determine how naturally the robot can interact with people and its environment. This report details recent research efforts focused on the structural design, auditory localization, and control system of a humanoid robot head, offering insights into a comprehensive approach that balances functionality, simplicity, and practical implementation.
The research addresses a fundamental challenge: how to design a humanoid robot head that can mimic essential human head movements while remaining mechanically feasible and cost-effective. While some humanoid robot heads are developed primarily to replicate facial expressions with dozens of degrees of freedom, this work takes a different path. It focuses on the neck and jaw movements necessary for natural interaction, resulting in a four-degree-of-freedom (4-DOF) configuration. This design philosophy reflects a pragmatic assessment of current technological capabilities and manufacturing constraints, particularly in the context of domestic research and development. The resulting humanoid robot head is lightweight, structurally sound, and capable of executing the basic motions required for tasks such as looking at a speaker, nodding in acknowledgment, and orienting toward sounds.
-
Background and Motivation for Humanoid Robot Head Development
The development of humanoid robots has progressed significantly over the past several decades. Early efforts in Japan, the United States, and Europe laid the groundwork for bipedal locomotion and basic human-robot interaction. Pioneering work by researchers such as Ichiro Kato at Waseda University in the late 1960s and 1970s produced some of the first bipedal walking machines, which eventually led to more sophisticated humanoid robot platforms. Honda’s P-series and ASIMO demonstrated remarkable advances in dynamic walking and balance, while Sony’s SDR-3X and later QRIO showcased the potential for entertainment and communication-oriented humanoid robots. In China, institutions such as the National University of Defense Technology and Beijing Institute of Technology have made significant strides, developing bipedal robots and humanoid robot platforms capable of complex movements, including martial arts demonstrations.
However, the head of a humanoid robot presents unique challenges and opportunities. Unlike the arms or legs, the head must integrate multiple sensory systems—vision, hearing, smell, and sometimes touch—while also providing a platform for expressive communication. Many humanoid robot heads are designed with a high number of degrees of freedom to replicate facial expressions, using mechanisms for eyebrows, eyelids, lips, and jaw. For instance, the WE-3RV and WE-4 series from Waseda University feature over 20 degrees of freedom in the head and face, enabling a wide range of expressions and emotional displays. The Kismet robot at MIT used 15 degrees of freedom in the face alone to convey emotions. These sophisticated systems are invaluable for studying social human-robot interaction and cognitive robotics. Yet they are also complex, expensive, and challenging to control.
This research takes a step back and asks: what are the essential degrees of freedom needed for a humanoid robot head to perform the fundamental motions of the neck and jaw? The answer proposed here is a 4-DOF configuration: three degrees of freedom in the neck (allowing left-right rotation, up-down pitch, and side-to-side roll) and one degree of freedom in the jaw (for opening and closing the mouth). This design enables the humanoid robot to orient its head toward a sound source, nod, shake its head, and tilt it—actions that are critical for natural communication. By focusing on these core motions rather than facial expressions, the design achieves a high level of mechanical simplicity and robustness. The head can be manufactured with standard machining processes, assembled with relatively low-cost components, and controlled with a straightforward control system. This approach is particularly well-suited for humanoid robots intended for mobile applications where weight and power consumption are important considerations.
The motivation for this work is also rooted in the need for humanoid robots to interact with humans in a natural and intuitive manner. For a humanoid robot to be accepted as a helper or companion, it must be able to perceive and respond to human presence. Sound localization is a key component of this perceptual capability. When a person speaks, the humanoid robot should be able to turn its head toward the speaker, establishing eye contact and demonstrating attention. This requires not only the mechanical ability to move the head but also the sensory and algorithmic ability to determine the direction of a sound source. The research therefore integrates sound localization with the head’s motion control, creating a system that can autonomously orient toward auditory stimuli. This integrated approach represents a significant step toward more autonomous and socially aware humanoid robots.
-
Mechanical Structure Design of the Humanoid Robot Head
The mechanical structure of the humanoid robot head is the foundation upon which all other functions depend. It must provide the required degrees of freedom, support the weight of sensors and electronics, and be robust enough for everyday use. The design process began by evaluating several possible structural configurations for achieving the required 4-DOF motion. These configurations included mechanisms with offset axes, intersecting axes, and remote actuation via linkages or belts. Each option was assessed for its complexity, manufacturability, and reliability. A key consideration was the routing of motor and sensor wires; configurations that allowed wires to twist or bind were rejected. Another consideration was the torque requirements on the motors. Configurations that placed excessive bending moments on a single actuator were deemed impractical. The selected design uses a universal joint principle combined with a rotational joint, creating a compact and efficient mechanism.
The humanoid robot head design features a neck structure with three degrees of freedom and a jaw mechanism with one degree of freedom. The neck’s three degrees of freedom are achieved through a modified universal joint arrangement. Specifically, the design allows for left-right rotation (yaw), up-down pitch, and side-to-side roll. The jaw’s single degree of freedom enables opening and closing, which can be used for simple expressions or to simulate speech. The head is constructed primarily from aluminum alloy, chosen for its favorable strength-to-weight ratio. The overall dimensions are approximately 140 mm in width, 180 mm in depth, and 380 mm in height, closely resembling the proportions of a human head. The total weight, including two small cameras, structural components, and the outer shell, is about 2 kilograms. This lightweight construction reduces the torque requirements on the neck motors and makes the humanoid robot head suitable for mounting on a mobile platform.
To determine the range of motion for each degree of freedom, the design specifies the following: the primary pitch (nodding) motion ranges from -45 degrees to +45 degrees, and the secondary pitch motion also ranges from -45 degrees to +45 degrees. The yaw (turning left and right) motion ranges from -90 degrees to +90 degrees. The lateral roll motion ranges from -45 degrees to +45 degrees. These ranges are presented in Table 1. The maximum torque required for each joint was estimated based on the weight of the head and the moment arm, yielding a requirement of approximately 2 N·m. This torque level is within the capability of small, high-performance DC servo motors when paired with a suitable reduction gear. The design uses direct drive for each joint, meaning the motor is directly coupled to the joint axis without belts or pulleys. This approach reduces complexity, minimizes backlash, and improves control precision. The trade-off is that the motors must be sized to provide the necessary torque at the relatively low speeds required for head movements.
The actuator selected for the humanoid robot head is the Maxon RE series DC servo motor. This motor features a high-energy rare-earth magnet and an ironless cup rotor, which provides high performance in a small package with low inertia. The specific model chosen is the RE 25-055-35 EBAZ0IA, which has a diameter of 25 mm and a rated power of 20 W. To achieve the required torque and speed, a planetary gearhead with a reduction ratio of 128:1 is employed. The gearhead has a diameter of 26 mm and can deliver a torque of up to 1.3 N·m. An encoder with 50 counts per revolution (and an index channel) is integrated into the motor assembly, providing position feedback for closed-loop control. The total length of the motor-gearhead-encoder assembly is approximately 118 mm. This compact unit is mounted directly at each joint, enabling a clean and space-efficient design. The use of servo motors rather than stepper motors was deliberate; servo motors offer higher torque density, better low-speed smoothness, and closed-loop control that prevents loss of synchronism under load. The control system, described later, further exploits these advantages.
The assembled humanoid robot head was tested for its mechanical integrity and range of motion. The universal joint design proved to be reliable and allowed smooth movement in all axes. The wiring for the motors and sensors was routed to avoid interference and twisting during head motion. The lightweight aluminum construction kept the overall weight within the design target, making the head easy to integrate with a robotic torso. Compared to more complex humanoid robot heads developed elsewhere, this design sacrifices facial expressiveness for mechanical simplicity and robustness. However, for the intended applications—such as a mobile robot that needs to look at and interact with people—the 4-DOF configuration is sufficient. It allows the humanoid robot to make eye contact, follow a conversation partner, and express basic attention and agreement through head nods. The design thus represents a practical compromise between human-like appearance and engineering feasibility.
Table 1: Range of motion for the humanoid robot head’s degrees of freedom Degree of Freedom Motion Type Range (degrees) Primary pitch Nodding (up-down) -45 to +45 Secondary pitch Nodding (up-down) -45 to +45 Yaw Turning left-right -90 to +90 Lateral roll Side-to-side tilt -45 to +45 Jaw Opening-closing Not explicitly stated, but single DOF -
Sound Localization System for the Humanoid Robot
For a humanoid robot to interact naturally with humans, it must be able to locate sound sources in its environment. Human hearing achieves this through a combination of interaural time differences (ITD), interaural level differences (ILD), and spectral cues. The humanoid robot head developed in this research takes a simplified but effective approach to sound localization, using two microphones, a standard computer sound card, and a cross-correlation algorithm to estimate the direction of a sound source. The system is designed to be low-cost, easy to implement, and capable of operating in real time. While more sophisticated systems use microphone arrays or binaural cues combined with head-related transfer functions, the two-microphone cross-correlation method offers a good balance between performance and complexity for a humanoid robot with limited computational resources.
The sound localization system consists of two electret condenser microphones, two sound cards (one integrated on the motherboard and one external USB sound card), and software running on a standard PC. The microphones are omnidirectional and have a relatively flat frequency response, making them suitable for capturing a wide range of sounds. The sound cards digitize the analog signals from the microphones at a sampling rate of 22,050 Hz with 8-bit resolution. The use of two independent sound cards allows the two channels to be sampled simultaneously, which is essential for accurate time-difference measurements. The software uses the DirectSound component of Microsoft DirectX to capture the audio streams in a streaming buffer, enabling real-time processing without the need to write data to disk. This approach minimizes latency and allows the system to process audio continuously.
The core of the sound localization algorithm is the cross-correlation function. For two discrete-time signals x(n) and y(n), the cross-correlation is computed as the sum of x(n+m)*y(n) over a window of samples. By finding the value of m that maximizes the cross-correlation, the system determines the time delay between the two microphone signals. This time delay, combined with the known distance between the microphones and the speed of sound, allows the angle of the sound source relative to the microphone array to be calculated. Specifically, if the microphones are separated by distance d, and the time delay is Δt, then the angle θ is given by θ = arcsin(c*Δt/d), where c is the speed of sound. This formula assumes a far-field sound source and a simplified model of the head as a sphere. The system implements this calculation and displays the estimated angle on a graphical user interface.
Several practical challenges were addressed during the development of the sound localization system. First, the microphones and sound cards introduce slight differences in gain and frequency response, which can affect the cross-correlation result. To mitigate this, the system includes a calibration step that normalizes the signals. Second, reverberation and background noise can create spurious peaks in the cross-correlation function. The system uses a windowing technique (e.g., a rectangular window of 50 samples) and selects the maximum correlation peak within a plausible range of time delays, reducing the impact of echoes. Third, the limited sampling rate (22,050 Hz) means that the time resolution is about 45 microseconds, which corresponds to an angular resolution of roughly 5 degrees for a microphone separation of 0.2 meters. While this is not as precise as human hearing, it is sufficient for the humanoid robot to turn its head toward a speaker and maintain a rough orientation. Experimental results showed that the system could reliably localize a sound source within a few degrees of its true position in a quiet room.
Despite its success, the sound localization system has limitations. The two-microphone configuration suffers from front-back ambiguity: a sound source directly in front of the microphones produces the same time delay as a source directly behind them. To resolve this, the system could be augmented with a third microphone or with visual information. The research acknowledges this limitation and suggests that future work could integrate the sound localization with the humanoid robot’s vision system to disambiguate the direction. Another limitation is the sensitivity to noise and reverberation, which is a common challenge in real-world environments. The use of more advanced algorithms, such as the generalized cross-correlation with phase transform (GCC-PHAT), could improve robustness. However, the current implementation demonstrates that a simple cross-correlation approach can provide useful directional information for a humanoid robot at a very low cost. The integration of this sound localization capability with the head’s motion control is a key step toward autonomous auditory orienting behavior.
-
Control System Implementation for the Humanoid Robot Head
The control system of the humanoid robot head is responsible for translating commanded angles into precise motor movements. It must handle four servo motors, each with its own encoder, and provide smooth, accurate motion while rejecting disturbances. The system is built around a PMAC (Programmable Multi-Axis Controller) from Delta Tau, specifically the PMAC2A-PC/104 model. This controller is a powerful and flexible motion control card that can control up to eight axes of various motor types, including DC servo motors. It features a Motorola DSP56300 series digital signal processor running at 40 MHz, 128 KB of SRAM, and 512 KB of flash memory. The card communicates with a host PC via a serial port or a PC/104 bus and can operate either as a standalone controller or as a slave to a PC. The control algorithms implemented in the PMAC include PID with velocity feedforward and notch filters, which are well-suited for precise servo control.
For this application, the PMAC is configured to control four Maxon DC servo motors through four Copely Acelnet Micro Panel servo amplifiers. Each amplifier receives a ±10 V analog command signal from the PMAC and drives the motor with a current proportional to the command. The amplifier also processes the encoder feedback signals and provides them to the PMAC. Because the Maxon encoders output single-ended TTL signals while the Copely amplifiers require differential signals, a signal adapter board was designed using an MC3487N quad differential line driver. This board converts the single-ended encoder signals to differential signals, ensuring reliable transmission over the cable between the motor and the amplifier. The wiring between the PMAC, amplifiers, motors, and encoders was carefully routed to minimize electrical noise and ensure proper grounding.
The PMAC control card was configured and tuned using the manufacturer’s PWIN32 software. Key parameters such as the PWM frequency, phase clock frequency, servo clock frequency, and DAC output range were set to match the motors and amplifiers. The servo loop was closed using the encoder feedback, and the PID gains were tuned to achieve fast response without overshoot or oscillation. The proportional gain (I130) was set to 200, and the derivative gain (I131) was adjusted to eliminate low-frequency oscillations. The integral gain was set to a small value to reduce steady-state error without causing instability. The resulting control loop provided smooth motion and accurate positioning. The PMAC also supports trajectory generation with S-curve acceleration, which reduces mechanical jerk and prolongs the life of the gearheads. The maximum speed and acceleration were set conservatively to ensure safe operation during testing.
The host software for the humanoid robot head was developed in Microsoft Visual C++ 6.0 using the MFC framework. The software provides a graphical user interface that allows the operator to command the head to specific angles, view the current position, and monitor the status of each axis. Communication with the PMAC is achieved through the PTalkDT ActiveX control, which provides a simple and reliable way to send commands and receive responses. The PTalkDT control handles the low-level protocol details, allowing the developer to focus on the application logic. The software includes functions for homing each axis on startup, setting the zero position, and moving to absolute or relative angles. The angle commands are converted to encoder counts using a calibrated scale factor. For the Maxon motors with 50-line encoders and 128:1 gear reduction, one revolution of the output shaft corresponds to 50 * 4 * 128 = 25,600 counts. However, after calibration, the actual count per revolution was found to be 25,600, and this value is used in the software to ensure accurate positioning.
Extensive testing was conducted to validate the control system. The humanoid robot head was commanded to move through its full range of motion in each axis, and the actual position was compared with the commanded position. The results showed that the system could achieve positional accuracy within a few encoder counts, corresponding to less than 0.1 degrees at the output shaft. The repeatability was also excellent, with the head returning to the same position after multiple movements. The maximum speed was limited to 60 revolutions per minute at the output shaft, which is sufficient for natural-looking head movements. The control system also demonstrated the ability to follow sinusoidal trajectories smoothly, which is important for tracking moving sound sources or people. The integration of the sound localization system with the control system was tested by having the humanoid robot turn its head toward a loudspeaker emitting a tone. The system successfully oriented the head within 0.5 seconds of the sound onset, demonstrating the feasibility of auditory-guided motion.
-
Experimental Results, Challenges, and Future Directions
The humanoid robot head system was assembled and tested in the laboratory. The mechanical structure proved to be robust and reliable, with no signs of wear or backlash after extended operation. The sound localization system was able to determine the direction of a sound source with an average error of less than 10 degrees in a quiet room, which is adequate for the robot to turn toward a speaker. The control system achieved precise position control, with the head moving smoothly to commanded angles. The integrated system was able to perform a simple auditory orienting task: when a sound was presented from a particular direction, the humanoid robot turned its head toward that direction within a fraction of a second. This demonstrated the successful integration of sensing, control, and actuation. The overall system represents a practical and cost-effective approach to building a humanoid robot head that can interact with humans through head movement and sound localization.
Several challenges were encountered during the development process. One of the main challenges was the tuning of the servo control loops. Because the motors are directly coupled to the joints without any compliance, the system is stiff, and improper PID gains can lead to oscillation or instability. The use of the PMAC’s built-in filtering and the systematic tuning procedure helped to overcome this. Another challenge was the sound localization in the presence of background noise. The cross-correlation method is sensitive to noise, and in a typical laboratory environment with computers and air conditioning, the signal-to-noise ratio was not always favorable. The use of a bandpass filter and a windowed correlation function helped to improve robustness. A third challenge was the integration of the two sound cards. The DirectSound API requires careful handling of buffer notifications to avoid dropouts, and the software was designed to use a double-buffering scheme with event-driven callbacks to ensure continuous capture.
Despite these challenges, the system met its design goals. The humanoid robot head is lightweight, has a sufficient range of motion, and can localize sounds and orient toward them. It is suitable for mounting on a mobile robot platform, where it can serve as a platform for further research in human-robot interaction. The lessons learned from this project can inform future designs. For example, the addition of a third microphone or a microphone array could resolve the front-back ambiguity and improve localization accuracy. The addition of more degrees of freedom in the face could enable facial expressions, though this would increase complexity and cost. The control system could be extended to include impedance control or force control, which would be useful if the head comes into contact with people or objects. The software architecture could be expanded to include higher-level behaviors, such as tracking a moving speaker using both sound and vision.
Looking forward, the research team plans to address the limitations identified in this study. One priority is to improve the accuracy of the sound localization system by using a more sophisticated algorithm, such as the generalized cross-correlation with phase transform (GCC-PHAT), which is more robust to reverberation. Another priority is to integrate the sound localization with a vision system, so that the humanoid robot can use both auditory and visual cues to locate and track people. This multi-modal approach would greatly enhance the robot’s ability to interact in real-world environments. Additionally, the team intends to develop a more advanced control strategy that allows the head to move in a more human-like manner, with coordinated movements of the neck and jaw. This could involve the use of minimum-jerk trajectories or central pattern generators. Finally, the team plans to test the humanoid robot head in a more realistic setting, such as a busy office or a home, to evaluate its performance under realistic noise and lighting conditions. These future directions will build on the solid foundation established by the current research.
In conclusion, the development of a humanoid robot head with integrated structure, sound localization, and control systems represents a significant contribution to the field of humanoid robotics. The 4-DOF design offers a practical balance between functionality and simplicity, making it accessible for research and education. The sound localization system demonstrates that low-cost components and straightforward algorithms can achieve useful directional hearing. The control system showcases the capabilities of modern motion control hardware and software in achieving precise, smooth motion. Together, these subsystems form a humanoid robot head that can see, hear, and move—a small but important step toward robots that can interact with humans in a natural and intuitive way. As humanoid robot technology continues to advance, the lessons from this project will help guide the development of more capable and affordable systems for the benefit of society.
The research described here was supported by a university’s Phase II 211 Project and involved collaboration among faculty and students. The work has been presented at academic conferences and has contributed to the training of graduate students. The humanoid robot head prototype continues to be used as a testbed for further research in sound localization, motion control, and human-robot interaction. Its modular design allows for easy upgrades and modifications, ensuring that it remains a valuable platform for years to come.