OpenWAM is a research stack for world–action models. It combines modular training, deployment and evaluation with controlled studies of how video priors and embodied pretraining affect robot control. Its OpenWAM-α model was pretrained on 518.5 million frames of egocentric and robot experience.

Overview

Model type
Action policy
Architecture
WAM
Listed versions
OpenWAM-α · LIBERO

Model history

November 2026

CoachSI post-training of OpenWAM on RoboCasa365

New capabilityOfficial announcement

Researchers at the University of Illinois Chicago collect rollouts from the task start. Frozen Qwen3.8-27B chooses an instruction at each OpenWAM action chunk. The researchers relabel successful trajectories with the original task instruction and fine-tune OpenWAM. Evaluation uses separate kitchen scenes and only the task instruction.

This announcement

Across 18 long-horizon tasks, 20 unseen scenes and three trials per scene give 1,080 episodes. Unassisted success rises from 47.4% to 56.8% with image-only coaching at collection or 59.3% with simulator-state coaching. Training uses 3,000 steps, batch size 16, 80% human demonstrations and 20% comparison data. Only video-backbone self-attention is updated, and the action backbone stays frozen. OpenWAM scores 79.2% on atomic tasks. The human-demonstration and self-collected-success baselines reach 49.4% and 51.9%. The variable 20% of the data has a matched budget of 530 samples per task and comes from extra human demonstrations, self-collected successes or coached successes. The image-only coach receives three camera views. Oracle next-step prompts yield 4.2 times the original instruction’s subgoal completion rate within 320 steps. The two base policies contribute a total of 500 stalled states from 15 tasks. A stall means 320 steps without subgoal progress. Each restored state has six continuations per instruction source with matched seeds. Subgoal completion is measured within 320 steps and task success within 640.

The paper identifies a public OpenWAM checkpoint and gives no exact version, checkpoint identifier or revision. Training code is marked coming soon. Derived weights and their license, as well as the code license, remain unconfirmed.

OpenRAL revises its LIBERO rSkill v0.1.0 documentation for OpenWAM-α

New resourcesVendor reported

At 04:37:41–42 Taipei time on November 20, OpenRAL added its October 29 LIBERO simulation results to the v0.1.0 manifest and revised the model card to describe one absolute arm joint-position target plus an absolute gripper position. Only the manifest and model card changed.

Previously reported

The initial manifest had no benchmarks section. The initial model card described two arm joint-position targets plus a gripper position.

This announcement

OpenRAL ran 10 tasks per suite and one episode per task on Jetson AGX Thor with BF16, compilation disabled and seed 0. Spatial 10/10, Object 9/10, Goal 9/10, Long (libero_10) 10/10, or 38/40 (95%) overall.

Arm and gripper targets describe the rSkill adapter interface. OpenWAM natively predicts end-effector deltas, which the adapter converts into 7-D LIBERO OSC commands. The four eval JSON files are byte-identical to the initial release. Table 13 of the upstream paper reports 99.3% under the authors’ full LIBERO protocol.

View all important updates →

Papers & reports (1)

OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining

Paper available

arXiv: 2609.07398DOI: 10.48550/arxiv.2609.07398
Authors & affiliations (24)

Author order: Original order

  1. Yuran Wang

    National University of Singapore

  2. Siqiao Huang

    Tsinghua University

  3. Mingleyang Li

    Peking University

  4. Chenhao Zhang

    Peking University

  5. Jiaqi Liang

    Peking University

  6. Weiyang Jin

    The University of Hong Kong

  7. Yue Chen

    Peking University

  8. Xuemin Chi

    Zhejiang University

  9. Donghao Zhou

    The Chinese University of Hong Kong

First published
2026-09-07
Latest paper revision
v1 · 2026-09-07
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

Official resources

Submit a source or correction