trajs/parrot_60Hz: stamp sim_dt + decimation so the 60 Hz demos replay at 60 Hz
Metadata-only. Every episode payload is byte-identical to what is on main (verified per
episode: md5 of pickle.dumps(ep, protocol=4), the same fingerprint upload_hf.py dedups on,
before and after the rewrite). The only change is the top-level metadata of each bundle.
What was wrong
Each of these bundles carried control_freq_hz: 60.0 and not sim_dt / decimation.
The replayer (RoboVerse/scripts/advanced/collect_demo.py) needs the pair — it appliessim_params.dt = sim_dt and decimation to the scenario before the handler is built — and
without it silently falls back to the task-authored dt=0.002, decimation=5 (100 Hz).
So every 60 Hz demo here replayed 1.67x too fast: each recorded PD target was held for
10 ms instead of 16.667 ms.
Measured through the real replayer on trajs/parrot_60Hz/grasp_apple episode 0
(202 recorded frames = 3.35 s of recorded sim time):
| bundle metadata | dt / decimation used | per control step | steps | sim time |
|---|---|---|---|---|
| as published here | 0.002 / 5 | 10.000 ms (100 Hz) | 167 | 1.67 s |
| with this PR's stamp | 1/480 / 8 | 16.667 ms (60 Hz) | 177 | 2.95 s |
The values, and where they come from
Every bundle gets sim_dt = 0.0020833333333333333 (1/480) and decimation = 8, sosim_dt x decimation = 1/60 s exactly. Three independent lines of evidence, no guessing:
- Raw-take provenance. Every one of the 591 episodes across all 13 published 60 Hz
bundles was matched back to its rawparrot_simtake by actions-sequence hash (not
pickle md5 — the two trees serialize the same demo differently), and every matched take's
own metadata carries exactlycontrol_freq_hz: 60.0, sim_dt: 1/480, decimation: 8. - In-band measurement. Each recorded frame's
_mujoco_state.buffer[0]is the MuJoCo
clock. Across all 13 bundles the frame-to-frame spacing is 1/60 s to within 5e-13, and the
absolute timestamps lie on a 1/480 grid and not on the 0.002 grid a 100 Hz recording
would have produced. - The recorders' formula.
parrot_sim/teleop.py::_cadence_for_hz(mirrored innetsim/server.py) is the only producer of non-100 Hz takes:decimation = round(480/hz),dt = 1/(hz*decimation)-> 60 Hz gives (1/480, 8), and
100 Hz reproduces the authored (0.002, 5).
Scope
12 bundles, 543 episodes:
| bundle | episodes | control_freq_hz | sim_dt | decimation |
|---|---|---|---|---|
bimanual_lift_box |
81 | 60.0 | 1/480 | 8 |
close_cabinet_12 |
51 | 60.0 | 1/480 | 8 |
collect_fruit_bowl |
50 | 60.0 | 1/480 | 8 |
fruit_tray |
51 | 60.0 | 1/480 | 8 |
grasp_adjustable_wrench |
49 | 60.0 | 1/480 | 8 |
grasp_apple |
47 | 60.0 | 1/480 | 8 |
grasp_banana |
48 | 60.0 | 1/480 | 8 |
hammer_insert_nail |
5 | 60.0 | 1/480 | 8 |
hammer_insert_nail_wall |
2 | 60.0 | 1/480 | 8 |
open_microwave_2 |
59 | 60.0 | 1/480 | 8 |
pour_kettle_cup |
50 | 60.0 | 1/480 | 8 |
trash_bin |
50 | 60.0 | 1/480 | 8 |
| total | 543 |
Deliberately not included: trajs/parrot_60Hz/feast/tianji_wuji/ (48 episodes) — that is
the pour_bottle_cup data, which is being republished right now as the consolidatedtrajs/parrot_60Hz/pour_bottle_cup/tianji_wuji/ bundle in its own PR, with the same cadence
stamp (sim_dt = 1/480, decimation = 8 — derived independently from the same recorder
formula and confirmed against the same 48 raw takes). Two PRs writing one path would
conflict, so that bundle is left entirely to the consolidation PR. Note for the reviewer: if
that PR does not also retire the legacy trajs/parrot_60Hz/feast/ path, that path stays
under-stamped and wants the same one-line fix.
The 100 Hz bundles under trajs/parrot/ are untouched — their cadence matches the
task-authored sim params, so they replay correctly with or without the stamp.
Companion code PR
SomaStacks/WUJI_SimBench#118 fixes the packers (pack_for_hf.py / upload_hf.py) so newly
packed bundles carry the whole triple, and makes the replayer warn loudly instead of silently
defaulting when it meets an under-stamped bundle like these.