h2r-benchmarks: five humanoid box loco-manipulation benchmarks (Unitree G1)

The five benchmarks of the H2ROI summary table, as the exact files the evaluation runner reads (sha256-verified against the evaluation machine, not re-converted; per-env identical results re-checked from a clean download on 2026-11-09), plus the raw teammate source cases. Code: https://github.com/BFCmath/H2ROI (private; collaborator access), main, docs docs/EVALUATION.md, docs/CHECKPOINTS.md. Policies: daibfc/h2r-weights.

folder	benchmark	eval set	clips scored	eval-ready	MANIFEST.sha256 sha256
hoi2roi_floor/	HOI2ROI floor (feasible43)	hoi2roi_feasible43	43	132 files, 72 MB	d0b9c15776a5add8...
chois_floor/	CHOIS floor (chois121)	hoi2roi_chois121	121	134 files, 63 MB	95664031808681c6...
hm3d_floor/	HM3D floor (hm3d_hoifhli25)	hm3d_hoifhli25	25	35 files, 21 MB	d45945d9bce33548...
table/	Table -> floor (hoi2roi_table42, kept26)	hoi2roi_table42	26	52 files, 22 MB	2fd2b920201163d5...
putdown/	Floor -> table put-down (hoi2roi_putdown49)	hoi2roi_putdown49	49	60 files, 27 MB	a3d16a9d64075c0d...
assets/	box mesh + box constants + G1 robot description the sim / scorers read	-	-	43 files, 52 MB	3849a9945d7c2987...

Raw source cases (<name>/raw/): hoi2roi_floor 175 files / 0.04 GB, chois_floor 1573 files / 20.80 GB, hm3d_floor 225 files / 1.42 GB, table 448 files / 4.33 GB, putdown 809 files / 6.94 GB.

Each benchmark folder: eval/ (repo-relative paths the runner loads), subset/ (scored clip keys), meta/ (key map, supports), raw/ (source cases, own raw/MANIFEST.sha256), README.md (source, clip count, scoring), MANIFEST.sha256 (sha256 of eval/ subset/ meta/).

Download (one command)

From a checkout of the code repo (https://github.com/BFCmath/H2ROI main):

scripts_sonic_render/ckpt.sh pull bench:all     # every set + assets + the eval-time plan table, ~0.7 GB, sha256-checked
scripts_sonic_render/ckpt.sh pull squeeze_avg10_12k   # a policy from daibfc/h2r-weights (checkpoints/<name>/, config.yaml next to it)

bench:all writes every file to the repo-relative path the runner reads (data/eval_sets/..., gear_sonic/config/eval_sets/, assets/objects/, data/ood_motion/p23_boxmix_place_A/plans.npz); every byte is checked against the MANIFEST whose sha256 is pinned in gear_sonic/config/benchmarks.yaml. Single sets: bench:<folder> (+ bench:assets bench:plan_table); --raw also fetches raw/ into data/benchmarks/<folder>/raw/ (CHOIS ~21 GB). A machine with the pre-release HOI2ROI manifest.json (sha256 f400b207...) keeps it with a note (eval-equivalent: only provenance paths differ); --replace-superseded swaps in the Hub copy. Without the repo: huggingface-cli download daibfc/h2r-benchmarks --repo-type dataset --local-dir h2r-benchmarks, then copy <folder>/eval/* into the code checkout root.

What one evaluation reads (complete list)

input	where	per benchmark
reference motions: teacher_robot.pkl, teacher_objects.pkl, actor_robot.pkl, actor_objects.pkl + manifest.json	<folder>/eval/data/eval_sets/<set dir>/	all five
per-clip plans (lift / release flags + planned box orientation, 50 Hz)	<folder>/eval/data/eval_sets/plans/<set>_approx/	all five
set yaml: clip list, offsets, box_scale, scene cuboids (support / target_support), hold_tail_s	<folder>/eval/gear_sonic/config/eval_sets/<set>.yaml (also in the code repo)	all five
scored clip subset	<folder>/subset/*.txt (= the yaml clip list; table: kept26 of the 42 cases)	all five
box mesh largebox.obj (sim + scorers), box constants, G1 USD / MJCF / meshes	assets/eval/...	all five
box-estimator plan table plans.npz (box_aug_data_dir)	daibfc/h2r-weights data/p23_boxmix_place_A/plans.npz (sha256 608d4958...), pulled by bench:plan_table	floor, CHOIS, HM3D, put-down (table: plans_file=null)
steering matrices + presets, gate / protocol strings, scorers	code repo (scripts_sonic_render/steering/r15/cfg/A_nav150.npz, gear_sonic/config/steering/, scripts_sonic_render/eval/)	all five
policy checkpoint + its eval config.yaml	daibfc/h2r-weights	all five

Scene setup. The simulator builds a flat ground plane, the G1, the box (largebox, size randomized per episode by the protocol below; the motions were made for largebox x 0.7601; start pose = first frame of teacher_objects.pkl) and, for table and put-down only, one static collision cuboid per clip (its support / target_support in the set yaml: size, centre, yaw; the same values are stored in each clip's teacher_robot.pkl entry). No room geometry is simulated or scored on any benchmark; room meshes in raw/scenes/ (CHOIS, table, put-down) are for viewing. The goal is the reference final box position (no goal file). Each set README has a "Scene setup" section.

Run the benchmark (code repo, A100 / RTX 5090)

GPUS="0 1" scripts_sonic_render/run_fullbench.sh checkpoints/squeeze_avg10_12k/model_step_011000.pt out/sq_off1

(*_nav150_full = "steering full episode"; protocol strings in scripts_sonic_render/eval/bench_protocols.sh.)

Common protocol (benchmark v3, 2026-11-11)

10 repetitions per clip, eval seed 0, whole clip from the first frame. Box per repetition: height, width and depth each drawn independently, uniform [25, 30] cm (outer size), mass uniform [0.3, 0.7] kg, all drawn from (seed, clip, repetition); the box is the OMOMO largebox scaled along its own axes about its bottom face (the code derives an OBB-frame copy of largebox.obj for this; nothing extra to download). Planned-wrist proximity gate 0.55 m, A_nav150 steering (nav150_off1 = until interaction, nav150_full = whole episode). pickup = the box physically lifted (floor: above min(30 cm, 50 % of the reference lift) for 0.3 s with wrist contact, upright; table: physical lift v2 off the support, both wrists); success = pickup AND final box within 30 cm of the reference final box position AND the box released at the end (no robot contact in the final 0.3 s, resting on its support in any orientation; a box still sliding on the floor counts, on furniture <= 5 cm/s; put-down: box hull within 2 cm of the table top). Details per set in its README (the set READMEs describe an older protocol; the code repo's docs/EVALUATION.md is authoritative).

Older revisions (numbers not comparable): v2.1 (2026-11-10) = the same rules with one box scale U[0.75, 0.90] of largebox and a 1 cm put-down touch limit (BENCH_REV=2.1); v2 (2026-11-09) = scale U[0.75, 0.90], success without the release test (BENCH_REV=2). FIXED_BOX=1 = the fixed box (scale 0.7601, 0.5 kg) as a control.

License: non-commercial research only, see LICENSE_NOTE.md.