GreenWAM: A World Foundation Model for Robotics

AI summary

Overview

Research area: Robot learning — world models, vision-language-action (VLA) policies, and cross-embodiment foundation models.

Technical level: Advanced. The paper assumes familiarity with action-conditioned video/world models, policy learning, action spaces, distillation, and robotics benchmarks.

Scope: The paper introduces GreenWAM, a single robot world foundation model plus the training recipe behind it, and evaluates it both as a scene-and-action predictor and as a control policy across many robot bodies, benchmarks, and one real humanoid.

What This Paper Is About

Robotics research usually treats two capabilities separately: a world model that predicts how the scene will change in response to the robot's actions, and a policy that decides which actions to take. Models are typically built and judged for only one of these roles, and each robot platform records its actions in its own incompatible format, which prevents knowledge from being shared across machines.

The goal of this work is to test whether one robot world model can serve both roles — predictor and controller — across many different robot bodies, and to identify what design choices actually make that possible.

Key Contributions

1. GreenWAM, a world foundation model for robotics, presented together with the explicit recipe that produces it, rather than as a model alone.
2. A shared anatomical action space that describes 20 robot families in common joint coordinates, spanning bimanual arms through dexterous humanoids, so that heterogeneous hardware can be trained together.
3. A joint training scheme over those robot families in three modes: visual, action-conditioned, and world-action.
4. A policy formulation that splits an action expert off the generator, yielding policies derived from the same world model, plus a distilled single-step version for real-time control.

Main Findings

• World-model performance: As a world model, GreenWAM reaches top-tier trajectory, depth, and 3D accuracy on WorldArena 1.0, ahead of every world-action model and video generator it was compared against. Exact scores are not given in the abstract.

• Policy performance on simulation benchmarks: As a policy, it places among the top entries of the RoboTwin 2.0 leaderboard, achieves the highest zero-shot LIBERO-plus success rate among the compared methods, and is competitive with dedicated vision-language-action and world-action policies on CALVIN and LIBERO. The abstract does not report the underlying numbers.

• Real humanoid transfer: GreenWAM was adapted to a real humanoid robot, where it succeeded more often than policies adapted from π_{0.5} and from Cosmos 3 Edge.

• Real-time control via distillation: The model can be distilled into a single-step policy for real-time control, indicating the world model is usable as a deployed controller and not only as an offline predictor.

• Robot specialization helps both roles: Controlled experiments found that specializing across robot embodiments improves world prediction and control together, rather than trading one off against the other.

• When to split the action expert: Splitting an action expert off the generator at the start of adaptation is beneficial; the timing of the split matters.

• Retaining world-model modes preserves robustness: Keeping the world-model training modes active during policy adaptation preserves robustness to scene changes.

Methodology in Plain English

The researchers start by removing the format barrier between robots. Instead of letting every platform keep its own action convention, they define a shared anatomical action space that expresses 20 different robot families in common joint coordinates — everything from two-armed manipulators to dexterous humanoids — so one model can be trained on all of them at once.

The model is then trained jointly across those families in multiple modes: a purely visual mode, a mode conditioned on actions, and a joint world-action mode. This is what lets a single network both imagine future scenes and act.

To turn the world model into a controller, the team separates an "action expert" from the generator rather than training a policy from scratch, and they keep the world-model modes running during adaptation so the policy does not lose its understanding of how scenes change. Finally, they compress the policy into a single-step version so it can run fast enough for real-time control, and adapt the whole system to a physical humanoid. They also run controlled experiments that vary these design choices individually, to isolate what actually matters.

Why This Matters

Impact on research: The paper argues against the prevailing split between world-model research and policy research, and against per-robot action formats. If one model can genuinely serve both roles across 20 robot families, that reframes world models as general-purpose robot backbones rather than evaluation-only artifacts, and makes the shared action space itself a reusable contribution.

Real-world applications:

• Humanoid and bimanual manipulation, where a single model could be reused across different hardware instead of retrained per platform.
• Real-time robot control, enabled here by distilling the policy into a single-step form.
• Cross-embodiment transfer, letting knowledge learned on one robot family benefit another.
• Robust manipulation in changing environments, since keeping world-model modes during adaptation is reported to preserve robustness to scene changes.

Industry relevance: A shared action space across 20 robot families reduces the cost of porting a learned controller to new hardware, and single-step distillation addresses the latency requirements that keep large models out of deployed robots. The reported comparison against π_{0.5} and Cosmos 3 Edge on a real humanoid places this work directly against well-known commercial and academic baselines.

Future Directions

• Whether the shared anatomical action space extends beyond the 20 robot families described, and how new hardware is folded into it.
• Whether the model can be scaled to even more embodiments and more training modes without the recipe breaking down.
• Whether single-step distillation retains the accuracy advantages seen in the non-distilled model, and how far it can be pushed for tighter control loops.
• Whether the three controlled findings — specialization, early action-expert splitting, and retained world-model modes — generalize beyond the specific robot families and benchmarks tested here.
• Whether real-world evaluation can be broadened beyond a single humanoid to the wider range of bodies covered in simulation.

Target Audience

Robotics and embodied-AI researchers working on world models, vision-language-action policies, and cross-embodiment learning; engineers building general-purpose robot controllers or porting learned policies across hardware; and readers already comfortable with action-conditioned generative models and robot benchmark suites who want to understand how a single model can be trained to both predict and act.

Authors’ abstract

A robot needs a world model that predicts how its actions will change the scene and a policy that chooses those actions. Models are usually built and evaluated for only one of these roles, and every robot records its actions in its own format. We ask whether one robot world model can serve both roles across different bodies, and what makes this work. We present GreenWAM, a world foundation model for robotics, together with the recipe that produces it: (1) a shared anatomical action space that describes 20 robot families, from bimanual arms to dexterous humanoids, in common joint coordinates; (2) joint training on these families in visual, action-conditioned and world-action modes; and (3) policies that split an action expert off the generator. As a world model, GreenWAM attains top-tier trajectory, depth and 3D accuracy on WorldArena 1.0, ahead of every world-action model and video generator in our comparison. As a policy, it scores among the top entries of the RoboTwin 2.0 leaderboard, reaches the highest zero-shot LIBERO-plus success among the compared methods and is competitive with dedicated vision-language-action and world-action policies on CALVIN and LIBERO. We further adapt GreenWAM to a real humanoid, where it succeeds more often than policies adapted from $π_{0.5}$ and from Cosmos 3 Edge, and distill it into a single-step policy for real-time control. Controlled experiments yield three findings: robot specialization improves world prediction and control together; an action expert helps when split off at the start of adaptation; and keeping world-model modes during policy adaptation preserves robustness to scene changes.