By modelling all machines within a shared anatomical workspace, GreenWAM serves both as a predictive simulator and a control policy, even for a real humanoid.
A robot needs two things: a model that predicts how its actions will change the scene, and a policy that chooses those actions. These two bricks are usually designed separately, and each manufacturer records the movements in its own format. GreenWAM, presented on arXiv on October 13, 2026, attempts to bring the two roles together in a single foundation model, valid for very different bodies.
A three-step recipe
The authors describe 20 families of robots, from two-armed systems to humanoids with agile hands, in a common space of joints. The model is trained jointly in visual, action-conditioned and action-world modes, and then an expert module dedicated to action is detached from the generator to produce control policies.
As a world model, GreenWAM obtains according to the study the best results of its comparison on WorldArena 1.0 in trajectory, depth and 3D accuracy. As a policy, it ranks among the top of the RoboTwin 2.0 ranking and succeeds more often than policies derived from π0.5 or Cosmos 3 Edge once adapted to a real humanoid; a single-step distilled version allows real-time control. Experiments also suggest that specializing the model on robotics simultaneously improves prediction and control.
Read the full article on the original source
Go to the source →