CoRL 2025: Hierarchical RL Framework Adapts Quadruped Skills to Complex Terrain
breadli428 · x · 2026-08-27
Researchers from ETH Zurich presented a hierarchical reinforcement learning framework at CoRL 2025 that enhances quadruped robot mobility in complex terrains using motion priors. The approach pre-trains a low-level policy to imitate animal motions on flat ground, establishing a foundation of natural movement. A high-level, goal-conditioned policy then learns residual corrections to enable perceptive locomotion, obstacle avoidance, and navigation.
Key Contributions:
- Hierarchical Structure: Separates natural motion expression (low-level) from adaptive corrections (high-level).
- Generalization: Simulation and real-world experiments with the ANYmal robot demonstrate successful transfer of flat-terrain skills to rugged environments while preserving natural gait.
- Motion Regularization: Shows improved motion consistency compared to baselines trained without priors.
More from Embodied
- Microduck: $399 open-source biped robot supports sim-to-real RL training — mishig25 · 2026-08-27
- AI-Assisted Brain Surgery Demo — mtizard · 2026-08-27
- Chinese Firm Unveils Wheel-Leg Hybrid Robot Dog — aftahi_ai · 2026-08-27
- Worm-inspired soft robot with inflatable muscles crawls and avoids obstacles — tekbog · 2026-08-27
- Photon Matrix Builds 'Iron Dome' for Mosquitoes with AI and Lasers — TinfoilTricorn · 2026-08-27
- Humanoid hype cools with 45% market pullback from viral highs — eyishazyer · 2026-08-27