emopointer/supermimic-manipulation

SuperMimic manipulation checkpoints

Trained SuperMimic agents for three simulated ARX5 manipulation tasks from supermimic_manipulation. Each checkpoint is taken at 4,000 training iterations and holds the full agent: world model, actor, critic, observation normalizers and optimizer states. It can be evaluated directly or used to continue training.

Task	Held-out success	Per reset category 0 / 1 / 2 / 3	Training time
lemon	100.0%	100 / 100 / 100 / 100	1.8 h
box_in_bin	94.9%	97 / 91 / 95 / 97	1.9 h
mug_hang	96.9%	94 / 98 / 95 / 100	2.8 h

Success is the simulator's native task success over 256 complete episodes per task (64 per reset category) on the held-out reset banks in emopointer/supermimic-manipulation-reset-banks. Training used one NVIDIA A100-SXM4-80GB per run.

Files

<task>/model_004000.pt     # agent checkpoint
<task>/params/env.yaml     # training environment; evaluation checks it against the current one
<task>/params/agent.yaml   # SuperMimic hyperparameters
<task>/params/mdp.json     # world-model state, action and imagined-reward contract
<task>/params/runtime.json # training run settings and reset-bank provenance

Keep each params/ directory next to its checkpoint: clab eval policy reads <task>/params/env.yaml from there.

Usage

hf download emopointer/supermimic-manipulation --local-dir checkpoints
hf download emopointer/supermimic-manipulation-reset-banks --repo-type dataset --local-dir banks
clab eval policy --task box_in_bin --algorithm supermimic \
  --checkpoint checkpoints/box_in_bin/model_004000.pt \
  --reset-bank banks/box_in_bin_test.pt --episodes-per-type 64 --num-envs 64 \
  --seeds 100000 --split-label held_out --cuda-graph-mode plain \
  --output reports/box_in_bin.json

See the repository README for installation, the approach and the references (SuperMimic, arXiv:2610.28480; OmniReset, ICLR 2026).