quincyu/box-wam-standalone

Box-WAM standalone policies

Single-task Box-WAM policies of Box Trajectories as Compact State Representations for World Action Models (CoRL 2026): a frozen DINOv3 ViT-L/16 encoder and a 130M-parameter rectified-flow transformer that jointly denoises the future box trajectory of the manipulated object and a chunk of joint targets. Each policy was trained from 50 demonstrations. Checkpoints are self-contained (weights, training config, normalisation statistics) and load with BoxWAMPolicy.from_checkpoint from box-wam.

robotwin/<task>.pt      nine RoboTwin 2.0 tasks, 15-step chunks, one head camera, 14-dim joints

RoboTwin tasks: adjust_bottle, beat_block_hammer, click_alarmclock, grab_roller, handover_mic, lift_pot, open_laptop, pick_dual_bottles (two box slots), turn_switch.

Usage

huggingface-cli download quincyu/box-wam-standalone --local-dir checkpoints/standalone
# RoboTwin 2.0 evaluation (box-wam repository)
python scripts/evaluate_robotwin.py --config configs/eval/robotwin_boxwam.yaml \
    --checkpoint checkpoints/standalone/robotwin/open_laptop.pt --tasks open_laptop --seeds 1 2 3 --episodes 50

from boxwam.policy import BoxWAMPolicy
policy = BoxWAMPolicy.from_checkpoint("checkpoints/standalone/robotwin/open_laptop.pt", device="cuda:0")
out = policy.act([head_camera_rgb], joint_positions)   # out.actions (15, 14), out.boxes (15, 1, 8, 2)

Results (paper Table 1, success rate %)

RoboTwin 2.0 (3 seeds × 50 trials)	Avg	Adjust bottle	Beat hammer	Click alarm	Grab roller	Handover mic	Lift pot	Open laptop	Pick dual bottles	Turn switch
clean	78.3	97.7	67.7	95.0	100.0	65.3	84.3	80.0	62.0	52.3
randomized	37.8	88.7	29.3	29.3	78.3	0.0	22.7	40.3	16.7	34.7

Multi-task model: quincyu/pi05_robotwin_obb_eef · Dataset: quincyu/robotwin_eef_obb_27500 · Project page: https://box-wam.github.io/

Citation

@inproceedings{yu2026boxwam,
  title     = {Box Trajectories as Compact State Representations for World Action Models},
  author    = {Yu, Qinxi and Ding, Wenzhou and Wang, Gloria X. and Zhai, Albert J. and Wang, Shenlong},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}

The DINOv3 encoder weights are not included; they are downloaded through timm under the DINOv3 license.