ETH's LP-ACRL Trains ANYmal to Run 2.5 m/s Over Rough Terrain
ChongZzZhang · x · 2026-09-30
ETH Zurich's Robotic Systems Lab released LP-ACRL (Learning Progress-based Automatic Curriculum Reinforcement Learning) to scale quadruped locomotion to diverse rough terrain.
- Instead of a hand-designed difficulty schedule, it estimates learning progress online and shifts training toward tasks where the policy can still improve.
- It automatically samples terrain types, levels, and velocity commands without a predefined order.
- The learned controller transfers to an ANYmal D robot, achieving fast, stable locomotion up to 2.5 m/s linear and 3.0 rad/s angular velocity across stairs, slopes, gravel, and low-friction surfaces.
Published in IEEE Robotics and Automation Letters, 2026.
Related event: ETH's Automated Curriculum RL Gets ANYmal Robot to 2.5 m/s on Rough Terrain(3 posts)→
More from Embodied
- RoboPapers teases episode on steering pretrained robot policies with VLMs (VLS) — micoolcho · 2026-09-30
- Delta-0 robot demo: WUJI Hand brings dexterous manipulation into whole-body control — siyuanhuang95 · 2026-09-30
- Quadruped inspection robot gets spatiotemporal memory to track railway anomalies — Separate_Fill_7127 · 2026-09-30
- OpenRoboto Shift sells real-work data to fund robot training on Bittensor SN80 — markjeffrey · 2026-09-30
- OpenAI's Dots device rumored to require a Pro 200 subscription plan — testingcatalog · 2026-09-30
- Linus Ekenstam 3D prints a full skateboard quarter pipe at home — LinusEkenstam · 2026-09-30