1. Introduction

This report presents the results of the design, implementation, and evaluation of an omnidirectional treadmill-based simulation platform intended for training and assessing bipedal walking robots. The platform enables continuous locomotion in a constrained physical space while preserving natural gait dynamics, thereby facilitating robust training, testing, and benchmarking of locomotion control algorithms under diverse conditions.

In addition, this study incorporates the recently published work “MicCheck: Repurposing Off-the-Shelf Pin Microphones for Easy and Low-Cost Contact Sensing”, which introduces a lightweight and cost-effective approach for contact sensing. By leveraging acoustic sensing for detecting contact events, the proposed platform extends its sensing capabilities beyond conventional force and motion measurements.

2. System Design

The proposed system integrates a robotic locomotion interface with an omnidirectional treadmill, supporting unrestricted planar movement and real-time feedback. The design consists of the following components:

Mechanical Structure:

A low-inertia omnidirectional treadmill with a concave walking surface, enabling passive recentering of the robot. The surface is coated with low-friction materials to minimize resistance while ensuring stable foot-ground contact.

Locomotion Support and Stabilization:

A harness-based support mechanism combined with a passive gimbal structure allows vertical load compensation and prevents falls without constraining horizontal motion. This ensures safety during unstable gait phases and early-stage controller training.

Actuation and Drive System:

The treadmill employs multiple independently controlled omnidirectional rollers arranged beneath the walking surface. These rollers dynamically adjust velocities to counteract the robot’s motion, effectively keeping it centered while allowing arbitrary walking directions.

Sensing and State Estimation:

The platform integrates motion capture cameras, inertial measurement units (IMUs), and force sensors embedded in the treadmill surface. In addition, MicCheck, low-cost pin microphones are embedded near foot contact regions to capture high-frequency acoustic signals generated during foot-ground interactions. These signals are processed to detect contact timing, slippage, and subtle interaction dynamics that are difficult to capture with force sensors alone.

Control and Software Integration:

A real-time control framework synchronizes treadmill actuation with the robot’s locomotion. The system interfaces with robot controllers and simulation environments (e.g., reinforcement learning frameworks), enabling closed-loop training and evaluation. Adaptive control algorithms adjust treadmill response based on predicted robot motion to minimize tracking error. Acoustic sensing data from the microphone array is fused with conventional sensor data to enhance contact state estimation.

3. Implementation and Testing

The system was implemented and evaluated using a standard bipedal humanoid robot platform. Key performance metrics include:

Motion Compensation Accuracy:

The treadmill maintained the robot within a 5 cm radius of the center during continuous walking at speeds up to 1 m/s, demonstrating effective motion compensation across multiple directions.

Gait Stability and Repeatability:

Robots trained on the platform exhibited stable walking patterns with a reduction in lateral drift compared to fixed-ground experiments. Step variability decreased significantly, indicating improved gait consistency.

Contact Detection Enhancement:

By integrating the MicCheck-inspired sensing approach, foot-ground contact events were detected with higher temporal resolution. The system also enabled detection of micro-slip and impact characteristics, improving gait phase estimation and control responsiveness.

Training Efficiency:

When integrated with reinforcement learning algorithms, the platform reduced training time due to continuous locomotion without boundary interruptions.

4. Results and Discussion

The omnidirectional treadmill-based platform demonstrated substantial advantages for bipedal robot training and evaluation. Continuous walking without spatial constraints enabled longer and more diverse locomotion sequences, improving both controller robustness and generalization.

The integration of MicCheck-based acoustic sensing provided a novel and effective complement to traditional sensing modalities. Unlike force sensors, the microphone-based approach captures high-frequency contact dynamics, enabling more precise detection of initial contact, slip onset, and surface interaction characteristics. This multimodal sensing significantly enhanced the fidelity of gait phase estimation.

The platform proved particularly effective for data-driven approaches such as reinforcement learning, where uninterrupted motion and richer sensory input improved sample efficiency and policy robustness. The safety mechanisms further enabled aggressive experimentation without risking hardware damage.

However, some limitations remain. Acoustic sensing is sensitive to environmental noise and requires signal filtering and calibration. Additionally, discrepancies in treadmill surface properties compared to natural terrain may still influence locomotion dynamics, particularly at higher speeds, where minor tracking errors were observed.

5. Conclusion

This study successfully developed an omnidirectional treadmill-based simulation platform for bipedal walking robot training and evaluation, enhanced by the integration of a low-cost acoustic contact sensing approach by MicCheck. The system provides a controlled, safe, and efficient environment for continuous locomotion experiments, significantly improving training efficiency, gait stability, and contact awareness.

Future work will focus on improving noise-robust acoustic signal processing, enhancing high-speed response capabilities, and incorporating variable terrain simulation. Further integration of multimodal sensing, including vision and tactile feedback, will expand the platform’s applicability to more complex real-world locomotion scenarios.