feat(trajs): consolidate pour_bottle_cup — merge the 110 verified feast episodes (290 total) + publish the 60 Hz bundle with correct cadence

#34
by Everloom - opened
SomaStacks Inc org

Consolidates the pour_bottle_cup trajectory data under its canonical name and fixes two
metadata bugs. Do not merge without review. feast/ paths are deliberately left intact.

What changes, per path

Path Before After
trajs/parrot/pour_bottle_cup/tianji_wuji/trajectory.pkl.gz 180 eps 290 eps (180 kept at indices 0–179 + 110 merged from feast/ at 180–289)
trajs/parrot_60Hz/pour_bottle_cup/tianji_wuji/trajectory.pkl.gz (absent) 48 eps (copy of trajs/parrot_60Hz/feast/tianji_wuji, cadence metadata repaired)

Nothing is deleted and no episode is dropped or reordered: the existing 180 keep their exact
indices, so any reference to pour_bottle_cup/tianji_wuji episode i < 180 still resolves to
the same demo.

Episode provenance — every episode published here is replay-verified

  • 180 episodes already published under pour_bottle_cup/tianji_wuji (unchanged).
  • 110 episodes merged in from trajs/parrot/feast/tianji_wujifeast is a retired
    alias of pour_bottle_cup (same task, same checker), so these are the same task's demos
    under the collection-era directory name.
  • 48 episodes at 60 Hz from trajs/parrot_60Hz/feast/tianji_wuji.

All 338 pass state-replay against the live hand_track.pour_bottle_cup checker in MuJoCo
(the task's own env.step() success latch), re-run against these exact rebuilt files:
290/290 for the merged 100 Hz bundle and 48/48 for the 60 Hz bundle, 0 errors.

Dedup was done on the action sequence (the recorded dof_pos_target stream), never on the
pickle md5 — the two source trees serialize the same demo differently (_meta, front_table),
so md5 manufactures false uniques. Result: 290 episodes → 290 unique, 0 duplicates; the
180-set and the 110-set are fully disjoint. Independently corroborated by the raw takes:
parrot_sim/.../pour_bottle_cup/tianji_wuji/uploaded/ holds 338 takes = 286 legacy-100 Hz +
4 explicit-100 Hz (= the 290) + 48 at 60 Hz.

Metadata fixes

  1. num_episodes was wrong. The feast/tianji_wuji bundle declared num_episodes: 106
    while holding 110 episodes. The merged bundle declares the true count, 290.
  2. The 60 Hz bundle had no sim_dt/decimation. It carried only control_freq_hz: 60.0;
    pack_for_hf.py propagates that one field and drops the other two. Replay reads
    sim_dt/decimation to set its cadence, so every 60 Hz bundle silently replayed at the
    task-authored 100 Hz — each action advancing 10 ms instead of 16.7 ms. Now stamped
    sim_dt = 1/480 = 0.0020833333333333333, decimation = 8 (8 × 1/480 = 1/60 s exactly),
    which is parrot_sim/netsim/server.py::_cadence_for_hz(60) and matches the value recorded in
    all 48 raw 60 Hz takes. Verified live: replay now reports cadence=(0.00208333, 8).
    The pack_for_hf.py code fix is a separate PR in the code repo.
  3. Both bundles are stamped task: pour_bottle_cup / task_name: hand_track.pour_bottle_cup,
    and carry a merged_from provenance block (source path, episode count, destination index
    range, source sha256).

Every per-episode _meta field is preserved byte-for-byte (input_source,
input_source_provenance, operator_pose, operator_pose_source, init_arm_pose,
language_instruction). Serialized under numpy 1.26 (the delivery env pin).

feast/ is left intact — on purpose

trajs/parrot/feast/** and trajs/parrot_60Hz/feast/** are not touched. Deleting them is a
separate later phase, and soma_render/ still pins the 110-episode feast bundle to reproduce a
published video. Consumers can migrate to pour_bottle_cup/ at their own pace.

Known follow-ups (not in this PR)

  • trajs/parrot/pour_bottle_cup/openarm_wuji is not published here. Its 2 episodes could not
    be replay-verified: the openarm_wuji robot cfg declares its 40 finger joints as
    <side>_hand_finger<n>_joint<m> while the shipped MJCF compiles them as
    <side>_finger<n>_joint<m>, so the env raises KeyError at launch. That is a pre-existing
    code↔asset naming divergence, unrelated to this data change — the bundle stays unpublished
    until it is fixed and the episodes verify.
  • The merged bundle resolves init_arm_pose to P1 (episodes 286–289 carry that stamp; the
    rest are unstamped). This only sets the spawn pose for fresh collection/RL — replay restores
    each robot state from the pkl, and all 180 previously-B0 episodes still latch 180/180 here.
Everloom changed pull request status to merged

Sign up or log in to comment