Reproduce “Visual CPG-RL: Learning Central Pattern Generators for Visually-Guided Quadruped Locomotion” on OpenResearch

About this paper

Guillaume Bellegarda, Milad Shafiee, Auke Ijspeert

The Visual CPG-RL framework learns visually-guided quadruped locomotion by integrating Central Pattern Generators (CPGs) with Deep Reinforcement Learning (DRL) for robust navigation in cluttered environments. This hierarchical control system on a Unitree Go1 robot tolerates sensorimotor delays up to 60 ms and demonstrates that interoscillator coupling enhances robustness for visual tasks.

Key insights

• Explicit interoscillator coupling within CPGs, while not critical for 'blind' locomotion, significantly improves robustness for visually-guided navigation tasks, especially in cluttered environments.
• Memory-enabled LSTM policies generally achieve greater energy efficiency (lower Cost of Transport) and better velocity tracking compared to MLP policies in visually-guided locomotion.
• The CPG-DRL architecture inherently provides superior robustness to sensorimotor delays (up to 60-90 ms) compared to joint-space DRL baselines, offering a mechanism for stable locomotion despite latency.

Summary by alphaXiv.