WillXH/pushf-demos

Push-F demonstrations

173 teleoperated episodes (22,436 steps at 10 Hz, 37 minutes) of a point agent pushing an F-shaped block into a goal pose, collected for FlowPush. Push-F uses the same physics, workspace, agent, goal semantics, and observation format as Push-T, with an F block whose area is within 7.1% of the T's.

Format

pushf_teleop.tgz unpacks to pushf/pushf_teleop.zarr, in the Push-T replay schema (data/{state, action, img}, meta/episode_ends) plus per-episode QC metadata (seed, max_coverage, success, duration_s). Each step stores the pre-step pose ([agent_x, agent_y, block_x, block_y, block_angle]), the cursor target as the action, and a marker-free 96×96 RGB frame. It loads through FlowPush's Push-T zarr adapter; the repo's make data-pushf downloads and unpacks it.

Collection

One operator, browser-based teleoperation. Episodes save automatically when goal coverage crosses 0.90 and are discarded on a 400-step timeout or on request. Reset seeds start at 300000 and are never reused, so a seed identifies an episode. 167 episodes form the clean pool used in the paper's experiments; six above the 0.95 cutoff (five deliberate ceiling probes, one ordinary save) are kept but excluded from the committed subset manifests. Max coverage 0.900–0.945 (mean 0.912); episode length 46–255 steps (mean 125). The operator got faster over the campaign (episodes about 30% shorter by the last quarter at the same coverage).

License

MIT. The environment physics are vendored from real-stanford/diffusion_policy (MIT).