No More Blind Spots: Learning Vision-Based Omnidirectional Bipedal Locomotion for Challenging Terrain

Abstract

Effective bipedal locomotion in dynamic environments, such as cluttered indoor spaces or uneven terrain, requires agile and adaptive movement in all directions. This necessitates omnidirectional terrain sensing and a controller capable of processing such input. We present a learning framework for vision-based omnidirectional bipedal locomotion, enabling seamless movement using depth images. A key challenge is the high computational cost of rendering omnidirectional depth images in simulation, making traditional sim-to-real reinforcement learning (RL) impractical. Our method combines a robust blind controller with a teacher policy that supervises a vision-based student policy, trained on noise-augmented terrain data to avoid rendering costs during RL and ensure robustness. We also introduce a data augmentation technique for supervised student training, accelerating training by up to 10 times compared to conventional methods. Our framework is validated through simulation and real-world tests, demonstrating effective omnidirectional locomotion with minimal reliance on expensive rendering. This is, to the best of our knowledge, the first demonstration of vision-based omnidirectional bipedal locomotion, showcasing its adaptability to diverse terrains.

Approach

A key challenge in omnidirectional vision-based locomotion is the high computational cost of rendering omnidirectional depth images during reinforcement learning in simulation. Our approach addresses this through a three-stage training framework:

Stage 1: Blind Controller. We first train a robust blind locomotion policy using privileged terrain information. This controller learns to walk in all directions without visual input, providing a strong foundation for subsequent stages.

Stage 2: Teacher Policy. We train a teacher policy that uses ground-truth terrain heightmaps to supervise omnidirectional locomotion. The teacher builds on the blind controller and learns to modulate actions based on terrain geometry.

Stage 3: Vision-Based Student. Finally, we train a student policy that uses depth images instead of ground-truth heightmaps. The student is trained via supervised learning on noise-augmented terrain data, avoiding the need for expensive depth rendering during RL. We introduce a data augmentation technique that accelerates this training by up to 10x compared to conventional methods.

Experiments

We validate our framework through both simulation and real-world experiments on the Cassie bipedal robot. The robot is tested on challenging terrains including random stepping stones, stairs, and large step-ups (up to 0.5m, approximately 60% of leg length).

Results

Our method achieves omnidirectional walking at speeds up to 1 m/s using depth camera input. Key results include:

• 10x training acceleration compared to conventional rendering-based approaches
• Successful sim-to-real transfer on the Cassie robot
• Robust locomotion across diverse challenging terrains
• First demonstration of vision-based omnidirectional bipedal locomotion