ckwolfe/benchmarks-viz-tiles

Dexterous Manipulation Benchmark — Cross-Method Evaluation

Cross-method evaluation scaffolding for dexterous manipulation, comparing 4 methods × 4 hands × 2 datasets × 7 trajectories on identical hardware, scenes, and reference trajectories.

Tag: v0.7-mt-complete · Trained checkpoints in the private companion repo ckwolfe/benchmarks-trained-ckpts.

---

TL;DR

method	class	simulator	best metric	status
ManipTrans	closed-loop RL (residual)	IsaacGym	tracking err 4.0 cm @ oakink	21/28 cells (schunk retrain live)
DexMachina	closed-loop RL (PPO)	Genesis	ADD 0.24 m @ arctic	28/28 cells ✅
Spider	sampling (MJWP)	MuJoCo-Warp	tracking err 8.3 cm, succ 36% @ oakink	12/28 cells
Oracle (kinematic replay)	open-loop upper bound	MuJoCo	tracking err 0.0 (by definition)	28/28 videos, 20/28 evals

---

Results — averaged across all real eval cells

method	n cells	ADD ↓ (m)	tracking_err ↓ (m)	success ↑	cost / cell
ManipTrans (closed-loop, IG)	27	—	0.040	—	~2 s
DexMachina (closed-loop, Genesis)	225	0.244	—	0.191	~300 s
Spider (sampling, MJWP)	28	—	0.083	0.357	~600 s
Oracle (kinematic replay, MJ)	20	—	0.000	n/a	<5 s

All numbers are per-cell means; full per-(method,hand,traj) rows live in metrics/*.jsonl.

Interpretation notes for reviewers

• DM "5 seeds" = 5 training reps, shared RL seed=42. All 140 public DM checkpoints in staging/run_paths.json were trained with seed: 42 per their config.json. The 5-rep variance reflects rl_games's env-step + mini-epoch stochasticity, not a seed sweep.
• Open-loop vs closed-loop. Spider (sampling) and Oracle (kinematic replay) are fundamentally different from MT/DM (closed-loop learned controllers). Reading them as one leaderboard is misleading. Paper Table 1 splits into two sub-tables keyed on method class.
• Simulator confound. MT runs in IsaacGym, DM in Genesis, Spider in MuJoCo-Warp, Oracle in MuJoCo. Contact model, integrator, and timestep differ. Tracking-err is most comparable within-class; cross-class comparisons are illustrative only.
• Divergent-rollout guard. Rows where the rollout diverged (tracking_err_mean > tracking_err_max + 0.01 or add_mean > 5 m) are blanked to null at write time. Schema validator in shared/bench/schema.py enforces this on append.

---

Datasets × hands

dataset	trajs	hands covered
OakInk-v2	lift_board, pick_spoon_bowl, pour_tube, stir_beaker, uncap_alcohol_burner, unplug, wipe_board	allegro, inspire, schunk, xhand
Arctic	ketchup30, box30, mixer30, ketchup40, mixer40, notebook40, waffleiron40	allegro, inspire, schunk, xhand

Per-cell coverage (✓ real eval, · missing)

ManipTrans × OakInk

	lift_board	pick_spoon	pour_tube	stir_beaker	uncap	unplug	wipe_board
allegro	✓	✓	✓	✓	✓	✓	✓
inspire	✓	✓	✓	✓	✓	✓	✓
schunk	·	·	·	·	·	·	·
xhand	✓	✓	✓	✓	✓	✓	✓

Schunk retrain running (epoch ~550/3000, reward climbing). Arrival: v0.8-schunk-retrained.

DexMachina × Arctic

	ketchup30	box30	mixer30	ketchup40	mixer40	notebook40	waffleiron40
allegro	✓	✓	✓	✓	✓	✓	✓
inspire	✓	✓	✓	✓	✓	✓	✓
schunk	✓	✓	✓	✓	✓	✓	✓
xhand	✓	✓	✓	✓	✓	✓	✓

---

Standardized videos

All videos are 720×480 @ 30 fps h264 crf 20, rendered through a shared MuJoCo scene with BENCH_CAMERA["front"] (pos=[0,-1.6,2.2], lookat=[0,-0.1,1.2], fov=30°). Files live under videos_std/.

bucket	count	prefix	description
Oracle (oakink)	28/28	std_*	kinematic replay through per-hand URDF
ManipTrans (oakink)	21/28	mt_*	captured rollout qpos replayed through same MJ scene
(MT × schunk)	0/7	mt_*	training in progress
Arctic	0/56	—	Spider arctic preprocess stage-2+ pending

Native (per-method) videos

Also published for cross-check: each method's own renderer (different camera, resolution, overlays). Under videos/:

• dexmachina_* (Genesis camera): 16 oakink demo-playback clips
• spider_* (MJX camera): 26 oakink cells
• ours_* (MuJoCo camera): 9 oakink cells

---

How to reproduce one cell

# Clone the benchmarking repo (not shipped here — the datasets repo only hosts artifacts).
git clone https://github.com/<you>/benchmarking && cd benchmarking
./scripts/run.sh maniptrans --hand xhand --dataset oakink_v2 --traj lift_board_bimanual --seed 42 --viz none --dev
# → appends MetricsRow to outputs/metrics/maniptrans.jsonl and (via bench_hooks.patch)
#   dumps qpos to outputs/mt_qpos/<run_id>.npz for post-hoc video rendering.

MUJOCO_GL=egl python scripts/render_standardized.py --source mt --hand xhand --traj lift_board
# → writes videos_std/mt_xhand_oakink_v2_lift_board_bimanual_seed42_default.mp4

Full sweep

./scripts/deploy_all.sh               # all 4 methods end-to-end, ~14 GPU-hr

Per-method deploy scripts under scripts/deploy_{spider,dexmachina,maniptrans,ours}.sh.

---

Files in this repo

metrics/
  maniptrans.jsonl     # 613 rows (27 real)
  dexmachina.jsonl     # 229 rows (225 real)
  spider.jsonl         # 293 rows (28 real — 1 divergent row stripped)
  ours.jsonl           # 20 rows (20 real)
videos_std/
  std_*.mp4            # 28 Oracle kinematic replays (oakink)
  mt_*.mp4             # 21 ManipTrans captured-qpos replays (oakink)
videos/
  dexmachina_*.mp4     # 16 DM native (oakink demo playback)
  spider_*.mp4         # 26 Spider native (MJX camera)
  ours_*.mp4           # 9 Oracle native (MuJoCo camera)
scripts/
  render_standardized.py   # MJ renderer shared across methods
patches/
  mt_bench_hooks.patch      # MT qpos-capture upstream patch (253 lines)
  mt_trajs_configs/         # 7 oakink YAMLs with upstream-compatible data_idx
STATUS.md                   # live coverage snapshot

---

Metric schema (shared/bench/schema.py::MetricsRow)

field	type	units	notes
run_id	str	—	canonical <method>_<hand>_<dataset>_<traj>_seed<N>_<warmstart>
success	bool	—	success_rate >= 0.5
success_rate	float	0–1	per-episode outcome mean
tracking_err_mean	Optional[float]	meters	position err vs reference; blanked if >1 m (divergent)
tracking_err_max	Optional[float]	meters	max per rollout; Spider uses quat-L2 legacy
add_mean	Optional[float]	meters	Average Distance of Displacement (DM canonical)
add_auc	Optional[float]	meters	AUC under threshold sweep 0–0.1 m
wallclock_s	float	seconds	host-side subprocess wall time
sim_steps	int	—	hard cap via BENCH_MAX_STEPS
upstream_commit	Optional[str]	git SHA	captured on host via BENCH_UPSTREAM_COMMIT env
sim_backend	str	—	one of isaacgym, genesis, mjwp, mujoco

Validator (at row-append time) rejects tracking_err_mean > tracking_err_max + 0.01 and add_mean > 5 m — those are the known divergent-rollout failure modes.

---

Limitations & what's not here

gap	why
DM × oakink training	140 public checkpoints are arctic-only; from-scratch is ~35 GPU-days per cell.
MT × schunk imitator	No public artifact; retrain in progress (v0.8 tag forthcoming).
Arctic standardized videos	Spider arctic preprocess stops at stage-1; no scene.xml for arctic cells.
Oracle real warmstart (Tara's retargeter)	Payload pending from an external collaborator; 55/56 stubs remain. Present arm uses kinematic IK fallback.
MT allegro/inspire video joint-name mapping	Hand-specific IG DOF ↔ MJ joint table needed (xhand works: 24/38 joints auto-matched).

---

Citation

Paper in preparation. If you reference this artifact in the meantime, please cite:

@misc{benchmarks-viz-tiles-2026,
  title  = {Dexterous Manipulation Benchmark — Cross-Method Evaluation Tiles},
  author = {C.K. Wolfe and T. Sadjadpour},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ckwolfe/benchmarks-viz-tiles},
  note   = {Tag v0.7-mt-complete}
}

Upstream methods retain their own licenses/attributions. See:

• ManipTrans — ManipTrans/ManipTrans; checkpoints: LiKailin/ManipTrans
• DexMachina — MandiZhao/dexmachina; 140 RL ckpts via upstream GDrive
• Spider — facebookresearch/spider; retargets: retarget/retarget_example