seasondish-try-1-raw

Raw recorded trajectories of MikasaSeasonDish-v0 — 1000 episodes, seeds 3000–4116 with gaps, 10 Hz.

Upload in progress. The .h5/.json pairs (~47 GiB) are still being transferred. Until this notice is removed, treat the episode set here as incomplete. The training-ready release is already complete.

For a training-ready version of the same 1000 episodes in LeRobot format (parquet + h264 video, 4.7 GiB instead of 47), see nurtayev-d/seasondish-try-1.

The task

A yellow ball hovers over one of two condiments on the counter, then disappears. The robot has to remember which condiment it marked, pick that one up, find the bowl, and hold the condiment tipped over it — leaving the other condiment untouched. The cue is gone by the time the robot can act, so the episode cannot be solved from the current observation alone.

Layout

seed_<N>/<timestamp>.h5      actions, agent state, both camera streams, full scene states
seed_<N>/<timestamp>.json    episode metadata: env config, seed, elapsed steps, success, provenance
seed_<N>.log                 per-episode recorder output (at the pool root, not inside the episode folder)
pool.env                     provenance of both passes: planning and cross-robot render
rerender.env                 provenance of the render pass alone
replay.jsonl                 one line per replayed episode — see below
jobs.txt                     the job list the pool was recorded from

Provenance

Two robots are involved, and that is deliberate.

1. Planning — commit 0285bdc, robot ds_fetch, planner season_dish_planner, state observations only. Seeds 3000–4099 gave 1083 successes out of 1100; a margin batch of seeds 4100–4199 added 99 out of 100.
2. Cross-robot render — commit 76136a9, robot fetch_cam224 (stock ManiSkill Fetch with 224×224 cameras). Each planned trajectory was replayed on the stock robot to produce the frames. 1072 of 1100 replays still reached the task's success predicate on the stock robot (97.5%), with joint-position divergence of 0.00095 rad median, 0.0043 at p90, 0.047 rad worst case.
3. Selection — the 1000 lowest seeds of those 1072 were kept; the other 72 were deleted.

replay.jsonl carries all 1100 replay attempts, not just the kept ones, with per-episode qpos_err, steps, wall time, and a problems list — so the 28 failures and the 72 dropped successes are documented rather than silently missing.

The render robot carries the ROBOT_FRONT_FACING_SIZE fix of 76136a9, which puts it where stock Fetch stands (0.8 m) — the same place the planner stood. Before that fix the identical replay kept only one episode in twelve.