Paper review · October 15, 2026
GreenWAM: A World Foundation Model for Robotics
The authors unified the actions of 20 robot types into common joint coordinates and trained jointly across several modes, producing a single robot foundation model that serves as both a world model and a policy.
Problem
Robots need both a world model that predicts how an action will change the scene and a policy that chooses actions. Models, however, are usually built and evaluated for only one of these roles. Each robot also records actions in a different format, which makes it hard to train across multiple robot embodiments together.
Method
Based on the abstract, the recipe has three parts. (1) Build an anatomically shared action space that represents 20 robot types, from bimanual robots to humanoids with dexterous hands, in common joint coordinates. (2) Train jointly in visual prediction, action-conditioned prediction and world-action modes. (3) Build the policy by separating an action expert from the generator. For real-time control, they also distilled it into a 1-step policy.
Results
- The authors report that on WorldArena 1.0, trajectory, depth and 3D accuracy surpassed all compared world-action models and video generation models
- As a policy, they report a top-tier position on the RoboTwin 2.0 leaderboard, the best zero-shot success rate among compared methods on LIBERO-plus, and performance comparable to dedicated VLA and world-action policies on CALVIN and LIBERO
- They report a higher success rate on a real humanoid than policies adapted from π0.5 and Cosmos 3 Edge
- From controlled experiments, they present three findings. Robot-specific training improves both prediction and control. Separating the action expert is effective when done early in adaptation. Keeping the world model mode during policy adaptation preserves robustness to scene changes.
Limitations
This analysis is based on the abstract. The full text was not available, so specific numbers, data scale, and the number and conditions of real-robot experiments could not be checked. Real-robot experiments were reported on one humanoid type.
Why it matters
It brings different robot embodiments into one action space and provides evidence that a single model can handle both prediction and control. It is a useful reference for data integration and training design in robot foundation models.
Key terms
- World model
- A model that takes the current scene and an action as input and predicts how the next scene will change.
- WAM(World-Action Model)
- An architecture that handles future scene prediction and action generation together in one model.
- Shared action space
- A representation that maps each robot's different action recording format onto common joint coordinates, so data from many robots can be used to train one model.
Original
Authors: Gennady Sidorov, Valerii Kornilov, Denis Grankin, Polina Fedotova, Aleksey Postnikov · Posted 2026-10-13 · License CC BY 4.0 · Based on the abstract
This review is an AI-written summary of the paper. Please treat the original paper as the reference.