pusht-simulator

Push-T on several physics engines with a measured-behaviour brain. Static pages and videos generated from the episode store; code and history at github.com/vizzzrohan/pusht-simulator; datasets pymunk-reference and superdex-reference.

• Contact-friction study — six stages, 25 seeds, 150 videos: what mu=0.5 between pusher and T does to the brain, and the brain that answers it
• 50-seed eval — pymunk-reference/rule vs pymunk-wall/contact2 on 50 fresh seeds, paired
• Write-up: contact friction — every number behind the study
• Session report: fidelity and behaviour — reference parity, human analysis, brain, store
• Human pushing analysis — 206 demonstrations, 15 figures, the measured behaviour the executor runs
• pymunk vs SuperDex, 5 seeds — same brain, two engines
• pymunk-reference vs pymunk-wall, 5 seeds — mu=0 vs mu=0.5, rule brain
• Seed 428 five ways
• All docs — convergence study, episode store, parity reports
• All episode videos — generated from the episode store

Engines: pymunk-reference (published gym-pusht, bit-exact), lewm (stable-worldmodel PushT-v1, bit-exact), pymunk-wall (published + contact friction), superdex-reference (published task on SuperDex contact), pymunk / superdex (SI re-derivations).