GreenWAM Unifies Robot World Modeling and Action Policies

Introduction

A capable robot needs more than a policy that maps observations to actions. It also needs a model of what will happen after an action is taken: objects may move, contacts may change, and the scene may evolve over time. In current systems, world models and action policies are often trained as separate components. The problem becomes harder when each robot describes its actions in a different format. GreenWAM addresses both issues with a single robotics foundation model intended to work as a world model and as a policy.

What the model changes

The most important part of the proposal is a shared representation for robot actions.

• A common anatomical action space: GreenWAM describes 20 robot families, ranging from bimanual arms to dexterous humanoids, using common joint coordinates. The goal is to make action knowledge less dependent on one specific robot body.
• Joint multimode training: The model is trained in visual, action-conditioned, and world-action settings. It must understand the current visual state, relate actions to future observations, and connect predicted consequences with executable behavior.
• A separate action expert: When used as a world model, the generator can represent possible futures under different actions. When used as a policy, an action expert is split from that generator to produce control outputs more directly. This retains a shared foundation while avoiding the need to use the full generative process at every control step.

Reported results

According to the paper summary, GreenWAM achieves top-tier trajectory, depth, and 3D accuracy on WorldArena 1.0, outperforming every world-action model and video generator included in the comparison. As a policy, it ranks among the leading entries on RoboTwin 2.0 and records the highest zero-shot LIBERO-plus success rate among the compared methods. On CALVIN and LIBERO, it is reported to be competitive with dedicated vision-language-action and world-action policies.

The authors also adapt the model to a real humanoid robot. The adapted GreenWAM policy succeeds more often than policies derived from π0.5 and Cosmos 3 Edge in the reported comparison. The team further distills the model into a single-step policy for real-time control. These claims should be read with the available evidence in mind: the supplied material does not include task-by-task scores, hardware details, training scale, or uncertainty estimates.

Why it matters

GreenWAM represents a shift from treating prediction and control as unrelated modules. If the same representation can support both, experience collected on one embodiment may become more useful to others, while world-model features could provide a stronger starting point for policy learning. The action expert also offers a practical route from a computationally expensive generative model to faster robot control.

There are still important open questions. A shared joint space does not automatically remove differences in sensors, kinematic limits, contact mechanics, latency, or data quality. Results from simulation may also fail to predict long-horizon reliability in the physical world. More details about training mixtures, compute requirements, failure cases, and reproducibility will be needed before GreenWAM can be judged as a broadly deployable robotics foundation model.

Source: Hugging Face Daily Papers