Train Microduck in the browser

Augmented Random Search (V2-t, with an ARS-inspired observation normalization: coordinates that have not varied are masked, and each normalized input is clipped to ±5) learns a linear correction (14 × 61 weights) added to the pretrained walking policy. Every episode is a real run of the whole-system FMU, from reset, for 3 s at 0.3 m/s forward command. The return is the trunk's forward progress, minus a small cost on the correction and a penalty for falling. Training happens in separate workers; the walking demo is unaffected.

Stop ends training at once: the running iteration is dropped and the learner keeps its last completed iteration. Training uses one worker. Save best policy saves the completed iteration with the highest evaluated return so far; training still continues from the current iteration, and Restore loads whichever policy was saved last.

Ready.

Current: iteration 0 · no completed iteration evaluated yet


Back to the simulator