Megadose The AI curve, compressed.

Auditing a Learned Simulator for Legged Robots: The Robotic World Model on ANYmal D

· ArXiv · AI/CL/LG ·
Better predictions did not reliably produce walking policies: 13 of 18 fine-tuned simulators yielded failures.

All 18 improved on the released model’s one-step and eight-step prediction accuracy. Training on the model’s own rollouts beat one-step training at 368 steps across all ten held-out episodes, despite scoring worse one step ahead. Ensemble disagreement ranked rollout errors well, but its one-sigma uncertainty estimates covered just 4.6% of outcomes, against 68% for a calibrated estimate. The audit covers one robot and one gait using simulated data, with no hardware tests.

ArXiv · AI/CL/LG's note

score 4

Categories: Research