Learning Bipedal Locomotion on Gear-Driven Humanoid Robot Using Foot-Mounted IMUs

TL;DR

This work tackles the sim-to-real gap in reinforcement learning for gear-driven humanoid robots lacking torque sensors by leveraging foot-mounted IMUs to rapidly stabilize locomotion across diverse terrains. It introduces symmetric data augmentation and random network distillation within a Legged Gym-based RL framework and employs a teacher–student training pipeline with fine-tuning to transfer policies to hardware. Hardware experiments on the EVAL-03 miniature humanoid demonstrate improved stability on non-rigid surfaces and during abrupt terrain transitions, though upward-step navigation remains a challenge. The approach offers a practical path toward robust, torque-free bipedal control for low-cost humanoids, with future work pointing to integrating additional sensing or lower-gain joint control to handle more complex terrains.

Abstract

Sim-to-real reinforcement learning (RL) for humanoid robots with high-gear ratio actuators remains challenging due to complex actuator dynamics and the absence of torque sensors. To address this, we propose a novel RL framework leveraging foot-mounted inertial measurement units (IMUs). Instead of pursuing detailed actuator modeling and system identification, we utilize foot-mounted IMU measurements to enhance rapid stabilization capabilities over challenging terrains. Additionally, we propose symmetric data augmentation dedicated to the proposed observation space and random network distillation to enhance bipedal locomotion learning over rough terrain. We validate our approach through hardware experiments on a miniature-sized humanoid EVAL-03 over a variety of environments. The experimental results demonstrate that our method improves rapid stabilization capabilities over non-rigid surfaces and sudden environmental transitions.

Figures (9)

• Figure 1: Upper: A photo (left) and kinematic model (right) of the gear-driven, miniature-sized humanoid robot EVAL-03. Lower: An overview of the proposed method. In the kinematic model, three IMUs mounted on the body, left foot, and right foot are illustrated as red spheres.
• Figure 2: Coordinate frame transformations in mirroring left-foot IMU observations to right-foot IMU observations for symmetric data augmentation.
• Figure 3: Teacher-student training with fine-tuning
• Figure 4: Plots of yaw velocities $w_{z}$ in tracking the turn command $w_{z, {\rm cmd}} = 0.5$ rad/s.
• Figure 5: Snapshots of EVAL-03 walking on floor using the foot-mounted IMU observations.