Event
FCTS: Training Quadrupedal Parkour Policies with Failure-Guided Terrain Curricula
1 reports · 1 sources
FieldAI proposes FCTS, a failure-conditioned terrain curriculum framework that automatically generates the next training terrain from a quadruped robot's local traversal failures; the failure location, termination mode, robot state, and surrounding terrain geometry jointly condition the generation process, and Monte Carlo tree search explores alternative curriculum trajectories under a fixed training budget. On terrain built from real construction-site scan data, FCTS improves the success rate on medium-difficulty curricula from 28.4% to 85.8% and on hard curricula from 14.9% to 44.7%, relative to the strongest parkour or LLM curriculum baseline. Real-robot experiments with the quadruped confirm that the learned policies transfer to the real world. The paper is currently under review.
Report timeline
- FCTS: Failure-Guided Curriculum Tree Search for Generating Terrain to Improve Quadruped Parkour Policies
FCTS is a failure-conditioned terrain curriculum framework that uses localized traversal failures to automatically generate the terrain a policy should train on next. Each failure's location, termination mode, robot state, and surrounding terrain geometry jointly condition the generation process, and Monte Carlo Tree Search explores alternative curriculum trajectories under a fixed training budget.
arXiv cs.RO ·