MAWAM

MAWAM adapts GR00T 2 for coordinated manipulation. It jointly predicts every robot’s future camera views and action chunks through cross-agent attention, while preserving separate control interfaces. The authors evaluated two Galaxea R1 Pro robots and teams of two to four simulated Panda arms.

View full model description

Papers & reports (1)

MAWAM: A Multi-Agent World-Action Model for Coordinated Manipulation

Paper available

arXiv: 2611.25105DOI: 10.48550/arxiv.2611.25105
Authors & affiliations (13)

Author order: Original order

  1. Xianzhe Fan

    NVIDIA · The University of Hong Kong

  2. Chi-Pin Huang

    NVIDIA

  3. Fu-En Yang

    NVIDIA

  4. Jing Wang

    NVIDIA

  5. Yecheng Wu

    NVIDIA · MIT

  6. Yu-Cheng Chou

    NVIDIA

  7. Johan Bjorck

    NVIDIA

  8. Linxi “Jim” Fan

    NVIDIA

  9. Jan Kautz

    NVIDIA

First published
2026-11-20
Latest paper revision
v1
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

How to access and use

Training code for this version is unconfirmed. Public code resources do not establish access to the exact training implementation. View verification notes

Training / fine-tuning code

SourceVerified on · 2026-11-20

Release information

SourceVerified on · 2026-11-20

Checks before deployment

Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Inference mode
Synchronized receding-horizon control, executing 8 of 24 steps before replanning; 15 Hz on Galaxea R1 Pro and 20 Hz in RoboFactory.
Fine-tuning requirements
Single-robot camera/controller warm-up, then joint training on synchronized demonstrations. RoboFactory uses 150 demonstrations per task.

Architecture and inputs / outputs

Model type
Action policy
Architecture
WAM · Transformer
Backbone
NVIDIA GR00T 2
Action generation
Flow matching
Input
Language / RGB images / Robot state / Video
Output
Galaxea R1 Pro: relative end-effector position and quaternion plus gripper opening per arm, 16 dimensions per robot. Simulated Panda: 17 dimensions per arm.

Scope and training

ee_pose · continuous

16 DoF · horizon 24 · relative

Output

Actions / Video

Prediction output

24-step action chunk per robot and eight future frames per private view, one frame every three control steps

Robot form

Dual-arm platform / Robotic arm

Task scope

Multiple tasks

Training

Imitation learning / Supervised fine-tuning

Lineage

Fine-tuned from: GR00T · GR00T 2 ↗

Related robots

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Full parameters and verification notes (model card)

Inputs include a language instruction, camera histories, robot states, embodiment metadata, roles, and view ownership. A shared simulation camera can supply read-only context. Each prediction covers 24 control steps and eight future frames per private view. Robots execute eight steps in sync and replan. The authors report 67.8% across three real tasks (30 trials each) and 90.3% across eight RoboFactory tasks (100 episodes each). Table 9 reports 2.76 billion trainable parameters, excluding frozen Cosmos-Reason1 and the video VAE.

Submit a source or correction