Quadruped robots have become adept at parkour over standard obstacles like steps and gaps, but real-world industrial and disaster-response sites contain irregular, compound geometry that typical training curricula overlook. A new preprint on arXiv addresses this gap by proposing FCTS, a training framework that generates terrain and uses failure-guided curriculum tree search to build more robust policies.
The abstract notes that canonical obstacles are handled well, but the irregular geometry of such environments is where agile locomotion matters most. By generating terrain and guiding the curriculum with failures, the method aims to expose policies to the kind of compound challenges they would face in the field.
Because this is a single preprint, there are no conflicting sources to compare. The full details of the method and results are not available in the abstract, so the claims here are limited to what the authors state in their summary.