ETH Zurich's Auto-Curriculum RL Gets ANYmal to 2.5 m/s on Rough Terrain With One Policy
breadli428 · x · 2026-09-30
ETH Zurich researchers Ziming Li, Chenhao Li, and Marco Hutter published a new IEEE RA-L (Aug 2026) paper on scaling rough terrain locomotion with automatic curriculum reinforcement learning.
Key results: after teacher–student distillation, a single policy transfers to the real ANYmal D quadruped, achieving:
- 3.0 m/s on flat terrain
- 2.5 m/s on challenging terrain
- Up to 3.0 rad/s angular velocity across all tested terrains
The approach automatically generates the terrain curriculum, avoiding hand-designed difficulty schedules while keeping one policy for real-world deployment.
Related event: ETH's Automated Curriculum RL Gets ANYmal Robot to 2.5 m/s on Rough Terrain(3 posts)→
More from Embodied
- Flourish home robot can be taught new tasks in 30 minutes via smartphone demos — philfung · 2026-09-30
- Friend founder pitches wearable AI pendant that could replace your phone and computer — jesselyu · 2026-09-30
- SLIP-VLA replaces iterative imagination with single-step latent for VLA policy learning — zhenjun_zhao · 2026-09-30
- ClearPath Open-Sources ROSCon 2026 Workshop: ROS 2 Middleware Tuning, DDS/Zenoh Benchmarks — ___Mufasaa · 2026-09-30
- SOLE-R1 accepted at NeurIPS: zero-shot reward prediction teaches robots 20+ new tasks from scratch — ___Mufasaa · 2026-09-30
- After 9 years in 16K homes, Matic argues robots are built in the wrong order — bigaiguy · 2026-09-30