jungwoopark01/DOMINO-eval-manifest

DOMINO fixed-episode evaluation manifests

Pre-screened episode sets for the DOMINO benchmark, demo_clean_dynamic, seed 0.

Why

DOMINO's default protocol screens candidate seeds online with an RRT expert planner until it finds 100 solvable episodes. That planner is stochastic, so two runs accept different episode sets and two policies end up measured on different physics.

A manifest pins the accepted seeds and each episode's dynamic-motion state -- target object, start/end position, velocity vector, duration, and an RNG snapshot (numpy, python, torch, torch_cuda). Every policy then sees identical episodes.

Producing these took roughly 12,414 expert-planner rollouts (~6 GPU-hours).

Use

git clone https://github.com/H-EmbodVis/DOMINO.git && cd DOMINO

# --include is important: without it, this repo's own README.md would
# overwrite DOMINO's README.md in your clone.
hf download <REPO> --repo-type dataset --include 'eval_manifest/*' --local-dir .
# (huggingface_hub < 0.34 spells the command `huggingface-cli download`)

python script/eval_policy.py --config <deploy_policy.yml> --overrides \
    --task_name adjust_bottle --task_config demo_clean_dynamic \
    --episode_manifest eval_manifest/adjust_bottle/demo_clean_dynamic/seed0.pkl

eval_policy.py validates task_name, task_config, embodiment, use_dynamic, dynamic_level and dynamic_coefficient against the manifest and refuses to run on a mismatch, so a manifest cannot silently be paired with the wrong configuration.

Provenance

• DOMINO commit: 22cd288865827cb932003c877cf2870fe5ca22e6
• manifest format_version: 1
• embodiment: ['aloha-agilex']
• use_dynamic=True, dynamic_level=1, dynamic_coefficient=0.1
• episodes per task: 100
• candidate seeds start at 100000

Caveats

• Python pickles carrying numpy RNG state. Load with a compatible numpy (DOMINO pins 1.26.4). EpisodeManifest.load rejects any other format_version.
• Physics replay is not bit-exact across machines; mm-level contact differences remain. What these remove is episode-set drift and initial-state drift, the dominant variance sources.
• The .json beside each .pkl is a human-readable summary. Evaluation reads only the .pkl.

Contents

35 tasks, 3,500 episodes, 12,414 candidate seeds screened (28% overall acceptance).

Sorted by acceptance rate -- low values mean the expert planner rarely solves that scene, so those manifests were the expensive ones to produce.

task	accepted	candidates	acceptance	seed range
put_object_cabinet	100	2004	5%	100030–102003
scan_object	100	1335	7%	100014–101334
place_bread_basket	100	1019	10%	100003–101018
dump_bin_bigbin	100	780	13%	100003–100779
move_playingcard_away	100	626	16%	100003–100625
place_shoe	100	614	16%	100000–100613
place_object_basket	100	419	24%	100000–100418
place_bread_skillet	100	321	31%	100002–100320
place_a2b_left	100	318	31%	100004–100317
shake_bottle_horizontally	100	313	32%	100004–100312
handover_block	100	299	33%	100002–100298
shake_bottle	100	299	33%	100000–100298
place_a2b_right	100	294	34%	100002–100293
put_bottles_dustbin	100	271	37%	100000–100270
place_can_basket	100	235	43%	100002–100234
place_fan	100	228	44%	100002–100227
beat_block_hammer	100	224	45%	100003–100223
hanging_mug	100	224	45%	100000–100223
handover_mic	100	223	45%	100000–100222
grab_roller	100	197	51%	100000–100196
move_can_pot	100	180	56%	100000–100179
rotate_qrcode	100	178	56%	100002–100177
adjust_bottle	100	172	58%	100000–100171
click_alarmclock	100	171	58%	100000–100170
stamp_seal	100	169	59%	100000–100168
move_pillbottle_pad	100	157	64%	100002–100156
place_mouse_pad	100	155	65%	100000–100154
press_stapler	100	155	65%	100000–100154
place_object_scale	100	137	73%	100000–100136
place_phone_stand	100	128	78%	100000–100127
place_container_plate	100	122	82%	100000–100121
move_stapler_pad	100	119	84%	100002–100118
place_empty_cup	100	117	85%	100000–100116
place_object_stand	100	109	92%	100000–100108
click_bell	100	102	98%	100000–100101

License

Apache-2.0, matching DOMINO. Please cite the DOMINO paper if you use these.