ckwolfe/dex-manip-benchmark

Dexterous Manipulation Benchmark — Cross-Method Video Grid

3 methods × 4 hands × 7 OakInk-v2 bimanual trajectories, all rendered under an identical 720×480 @ 30fps BENCH_CAMERA so the paper tile grid lines up frame-for-frame.

Metrics in ckwolfe/benchmarks-viz-tiles · trained checkpoints in ckwolfe/benchmarks-trained-ckpts.

---

Paper comparison table

method	class	simulator	n eval cells	ADD ↓ (m)	tracking_err ↓ (m)	success ↑	cost / cell	standardized videos
ManipTrans	closed-loop RL (residual)	IsaacGym	27	—	0.040	—	~2 s	21/28
DexMachina	closed-loop RL (PPO)	Genesis	225	0.244	—	0.191	~300 s	0/28
Spider	sampling (MJWP)	MuJoCo-Warp	28	—	0.083	0.357	~600 s	0/28

Reference trajectory videos (kinematic replay of the demonstrator under the same camera) are published separately as a visual quality reference — not a peer method. See the reference-only appendix section below if you need them.

---

Video grid — ManipTrans (OakInk-v2)

The paper's closed-loop-RL column. Videos are MT's captured rollout qpos replayed through the shared BENCH_CAMERA MuJoCo scene. Xhand joint names auto-match the MJ scene (24/38 joints); allegro / inspire currently fall back to a linear DOF mapping, which leaves the hand frozen (follow-up: per-hand name translation table).

allegro

lift_board	pick_spoon	pour_tube	stir_beaker	uncap	unplug	wipe_board

inspire

lift_board	pick_spoon	pour_tube	stir_beaker	uncap	unplug	wipe_board

schunk

lift_board	pick_spoon	pour_tube	stir_beaker	uncap	unplug	wipe_board
·	·	·	·	·	·	·

xhand

lift_board	pick_spoon	pour_tube	stir_beaker	uncap	unplug	wipe_board

---

Coverage summary

method	real eval cells	standardized videos
ManipTrans	oakink 27/28, arctic 0/28	21/28
DexMachina	arctic 225/28=8-rep, oakink 0/28	0/28 (native renders only)
Spider	oakink 28/28 sampled / 12/28 real row	0/28 (native renders only)

OakInk bimanual tasks: lift_board · pick_spoon_bowl · pour_tube · stir_beaker · uncap_alcohol_burner · unplug · wipe_board.

Arctic tasks: ketchup30 · box30 · mixer30 · ketchup40 · mixer40 · notebook40 · waffleiron40 (s01 subject).

---

Methods

ManipTrans — closed-loop residual-RL atop per-hand PPO imitator, IsaacGym. Paper · Upstream. This benchmark ships imitators for {allegro, inspire, xhand}; schunk retrain in progress.

DexMachina — rl_games PPO in Genesis, 140 arctic checkpoints from upstream. Paper · Upstream. OakInk not publicly released; from-scratch is ~35 GPU-days/cell.

Spider — sampling-based MPC on MuJoCo-Warp. Upstream. MJWP sampling complete 28/28 oakink; real-row MetricsRow emission partial.

---

Metric schema

shared/bench/schema.py::MetricsRow. Key fields:

field	units	notes
success_rate	0–1	per-episode outcome mean
tracking_err_mean	meters	position err vs reference; blanked if divergent (>1 m)
add_mean	meters	Average Distance of Displacement
sim_steps	—	hard cap via BENCH_MAX_STEPS
upstream_commit	git SHA	captured on host
sim_backend	str	isaacgym / genesis / mjwp

Row-append validator rejects tracking_err_mean > tracking_err_max + 0.01 (the divergent-rollout failure mode).

---

Reviewer notes

"5 seeds" = 5 training reps (shared seed = 42). All DM checkpoints share seed: 42 per each config.json. Variance reflects rl_games's env-step + mini-epoch stochasticity, not a seed sweep.

Open-loop vs closed-loop. Spider (sampling) runs in a fundamentally different regime from MT/DM (learned closed-loop). Paper Table 1 splits into two sub-tables keyed on METHOD_AXIS_CLASS in shared/bench/plot_style.py.

Simulator confound. MT=IsaacGym, DM=Genesis, Spider=MuJoCo-Warp. Contact model, integrator, and timestep differ — tracking-err is within-class comparable, cross-class illustrative only.

---

Known gaps

• DM × oakink training — public ckpts arctic-only; from-scratch is ~35 GPU-days/cell.
• MT × schunk imitator — retraining overnight.
• Arctic standardized videos — Spider arctic preprocess stops at stage-1.
• MT allegro/inspire video joint-name mapping — xhand 24/38 auto-matched; others need per-hand translation table.

---

Reference kinematic trajectories (not part of the method comparison)

For qualitative inspection of the reference demonstrations each method is trying to track, videos/std_*.mp4 contain the per-hand kinematic replay of the Spider-preprocessed trajectory under the same BENCH_CAMERA. These are not a 4th method — they are what the policies are being compared against.

---

Citation

@misc{dex-manip-benchmark-2026,
  title  = {Dexterous Manipulation Benchmark — Cross-Method Video Grid},
  author = {C.K. Wolfe and T. Sadjadpour},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ckwolfe/dex-manip-benchmark},
}