← Paper Digest

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

Xianzhe Fan, Chi-Pin Huang, Fu-En Yang, Jing Wang, Yecheng Wu, Yu-Cheng Chou, Johan Bjorck, Linxi "Jim" Fan, Jan Kautz, Yu-Chiang Frank Wang, Zhiding Yu, Hengshuang Zhao, Zhiqi Li

November 2026

Two real robots collecting laundry into a bag, folding a bath towel and handing over an umbrella (3x speed).

TL;DR MAWAM extends the single-robot world-action model GR00T 2 to robot teams, generating every robot's action chunk and future camera views in one joint sample. On two real dual-arm robots it averages 67.8% success over three collaborative tasks, 18.9 points above MA-VLA trained on the same demonstrations, and 90.3% on eight simulated tasks with two to four arms.

What they do

In team manipulation, each robot’s actions change what its teammates see and can do next. MAWAM treats this as a multi-agent world-modeling problem: one model jointly generates all robots’ action chunks and the future views from each robot’s own cameras, while every robot keeps its own controller.

How it works

  • Per-robot streams. Each robot’s private camera views, state and action tokens form one stream. Its views share an identity code: a vertex of a regular simplex placed on an extra agent axis of the video positional encoding. A separate role embedding marks its job, so two robots can share a role without merging their observations or controls.
  • Joint denoising. One Transformer, initialized from GR00T 2, attends densely across all agents and jointly denoises their 24-step action chunks and eight future frames per view, trained with flow matching.
  • Two-stage training. Warm-up adapts the backbone to each robot’s cameras and controller one robot at a time; team training then adds cross-agent attention, roles and randomized identities on synchronized demonstrations. At run time the team executes 8 of 24 steps in lockstep and replans.

Results

  • Two Galaxea R1 Pro robots, six private cameras, 30 trials per task: 67.8% mean success vs 48.9% for MA-VLA, 37.8% for CLS-DP and 28.9% for GauDP. Laundry-bag collection rises from 30.0% (MA-VLA) to 53.3%; without cross-agent attention the mean drops to 52.2%.
  • Eight RoboFactory tasks with two to four arms: 90.3% average vs 83.4% for MA-VLA, with the largest gains on Place Food (78.0% vs 59.0%) and Take Photo (97.0% vs 82.0%).
  • Ablations: without cross-agent attention the four-arm relay falls from 95.0% to 39.0%; without team-stage future-video supervision Place Food falls from 78.0% to 47.0%; without warm-up the relay falls to 34.0%.

Limitations

Experiments cover fixed teams of up to four agents. Changing team membership during execution, mixing robots with different embodiments and sensors, and human partners are left to future work. Joint inference takes 0.32 s per call on an H100 for the real two-robot setup and grows from 1.40 s to 1.85 s going from two to four simulated arms.

Videos

Simulated arm teams, with each arm's predicted future views shown next to the recorded ones.

Figures

From Figure 1: MAWAM generates every robot's future camera views and action chunk together. Left: two real robots, ea...
From Figure 1: MAWAM generates every robot's future camera views and action chunk together. Left: two real robots, each with three private cameras and no shared camera, hold a laundry bag open while one loads the garment. Right: two to four simulated arms; in Camera Alignment each arm's private view is predicted six control steps ahead.
From Figure 2: Warm-up adapts the pretrained single-robot backbone one robot at a time; team training then learns fro...
From Figure 2: Warm-up adapts the pretrained single-robot backbone one robot at a time; team training then learns from synchronized demonstrations. Each robot's stream carries its views, state, role and noisy action and future-view targets, marked with a simplex identity; one shared Transformer attends across all agents, and the robots execute the first 8 of 24 predicted steps together before replanning.
From Figure 3: For each arm in simulated Camera Alignment, the observed private view, MAWAM's prediction six control ...
From Figure 3: For each arm in simulated Camera Alignment, the observed private view, MAWAM's prediction six control steps ahead, and the executed view at that step. Right: predicted and executed views for Lift Barrier, Pass Shoe and Long Pipeline Delivery.
Abstract

Every move in a robot team changes the world in which its teammates must act. Coordinated manipulation therefore calls for predicting the team's actions together with the physical future they jointly create. We introduce MAWAM, a multi-agent world-action model that jointly generates each robot's future views and executable action chunks in a single sample. Built from a pretrained single-robot world-action backbone, MAWAM jointly denoises all robots' action and future-view tokens through cross-agent attention, while preserving a separate control interface for each robot. MAWAM achieves the highest success on all three real-robot tasks (laundry-bag collection, bath-towel folding, and umbrella handover), with 67.8% mean success, 18.9 percentage points above MA-VLA trained on the same demonstrations and cameras. In simulation, MAWAM reaches 90.3% mean success on eight tasks with teams of up to four arms. Ablations show that cross-agent attention and team-stage future-video supervision both improve coordinated control: without attention, a four-arm relay falls from 95.0% to 39.0%, and without this supervision, Place Food falls from 78.0% to 47.0%.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.