ResWAM-RoboTwin2 (Official RoboTwin 2.0 50-Task Checkpoint)

Official ResWAM multi-task bimanual policy checkpoint (OPTML-Group/ResWAM-RoboTwin2) trained across all 50 tasks of RoboTwin 2.0 without external robot pretraining, achieving 93.6% under Clean conditions, 94.0% under Domain-Randomized conditions, and 93.8% overall average success rate across 10,000 evaluation episodes (Table 2 of the paper).

• Paper: ResWAM: Learning What Changes, Not What Stays, for World-Action Models
• Authors: Chongyu Fan, Gaowen Liu, Ramana Rao Kompella, Jason Corso, Sijia Liu
• Affiliations: Michigan State University, Cisco, University of Michigan, IBM Research
• Project Page: https://chongyu-fan.netlify.app/posts/reswam/
• Code: https://github.com/OPTML-Group/ResWAM
• HF Collection: https://huggingface.co/collections/OPTML-Group/reswam-6af8906dc9b6f21539b30934

Abstract

World-action models (WAMs) augment vision-language-action policies (VLAs) with future prediction so that the policy can anticipate how its actions will affect the physical world. Existing WAMs supervise complete future observations. However, observations typically change little over an action chunk, so the future loss is dominated by copying static background content rather than learning task-relevant robot and object motion, limiting policy learning. Moreover, existing WAMs rely on large video generation models or separate world models, incurring heavy training and inference compute that hinders closed-loop deployment. To improve policy performance, we propose visual residual learning, supervising only the feature change from the current frame to the future frame in a frozen DINOv3 space. Compared with complete future features, this residual learning target yields a higher feature rank with richer and more diverse representations of scene changes, and spatially concentrates on the robot arm and manipulated objects rather than the background. To further improve computational efficiency, we present ResWAM, a residual world-action model with 2.3B parameters in total, including a lightweight 0.5B dual-branch DiT that jointly trains an action branch and a residual branch via conditional flow matching. Without robot pretraining (i.e., pretraining on external robot action trajectories beyond the benchmark demonstrations), ResWAM achieves top average success rates on LIBERO-Plus (81.1%), RoboTwin 2.0 (93.8%), and RoboDojo (9.40% success, 13.41 score) at the lowest training (2.7–4.0 TFLOPs) and inference compute (1.0–1.4 TFLOPs), and leads real-robot Open, Put, and Stack (80.0% average, +10.0% over π_0.5) at 92.7 ms latency.

Checkpoint Summary

Field	Value
Repository	OPTML-Group/ResWAM-RoboTwin2
Checkpoint file	model.pt
Parameters	2.3B (RynnBrain-1.1-2B 16-layer text backbone + 0.5B dual-branch DiT + frozen DINOv3-B/16)
Action horizon ($H$)	32 (14-D bimanual joint position + grippers; executes 16 steps with degree-3 B-spline smoothing)
Compute & Latency	4.0 TFLOPs train / 1.4 TFLOPs infer / 99.4 ms per action chunk (10 Euler steps, RTX A6000)

RoboTwin 2.0 Results (50 Tasks, 100 Trials / Task)

Clean (50 Tasks)	Randomized (50 Tasks)	Average	Latency
93.6%	94.0%	93.8%	99.4 ms

Files in this Repository

• model.pt: Full model weights and embedded codec metadata.
• codec_statistics.json: 14-D bimanual action and state z-score normalization statistics (ZScoreCodec.from_file("codec_statistics.json")).
• resolved_config.json: Complete resolved model, data, and evaluation configuration.

Download and Evaluation Usage

hf download OPTML-Group/ResWAM-RoboTwin2 --local-dir ./checkpoints/ResWAM-RoboTwin2

python scripts/eval/eval_robotwin2.py run \
    --checkpoint ./checkpoints/ResWAM-RoboTwin2/model.pt \
    --robotwin-root ./third_party/RoboTwin \
    --output-dir ./outputs/robotwin2/eval_p32e16 \
    --gpus 0,1,2,3

Citation

@article{fan2026reswam,
  title={ResWAM: Learning What Changes, Not What Stays, for World-Action Models},
  author={Fan, Chongyu and Liu, Gaowen and Kompella, Ramana Rao and Corso, Jason and Liu, Sijia},
  journal={arXiv preprint},
  year={2026}
}