One robot world model that also acts: GreenWAM maps 20 bodies onto 79 anatomical slots and hits 95.6% instruction success on a real humanoid

A single robot foundation model now both predicts how an action will change the scene and emits the action itself, sharing one action interface across different bodies. Robot world models and world-action models have usually been built and evaluated for only one of these roles, and the robot actions of Cosmos 3 Edge cover six embodiments as end-effector pose deltas with a gripper opening. On the 1,000 test episodes of WorldArena 1.0, GreenWAM reaches 56.0 trajectory, 98.2 depth and 93.4 3D accuracy, the highest among the compared world-action models and video generators; on the real Green humanoid it completes 95.6% of in-distribution instructions, against 88.9% for the Cosmos 3 Edge policy and 84.4% for π0.5.

Cross-body unification comes from one shared anatomical coordinate table: each body needs only a codec that selects, reorders and normalizes its native joints, after which all bodies share one action projector. With one action projection per embodiment domain, or with end-effector poses plus gripper state, the correspondence between the arms, torso, legs and fingers of different bodies stays implicit or absent. The 79 slots split into 12 leg, 23 body, 18 arm and 26 hand coordinates and cover 20 robot families; a validity mask removes absent or unresolved coordinates from the loss, so a joint a robot does not have never becomes a zero-motion target.

The policy splits an action expert off the same generator and keeps world-model modes during adaptation, and each choice raises success at a fixed adaptation budget. The comparison runs start from the same checkpoint and use the same recipe and step count, changing one design choice at a time. On RoboTwin 2.0, initializing from GreenWAM rather than Cosmos 3 Edge raises mean success by 5.4%; splitting off the action expert at the start of adaptation adds 3.3%; keeping forward dynamics and image-to-video during adaptation raises randomized-scene success by 8.0% with no significant change in clean scenes.

The demonstrated scope is the 20 two-armed robot families in the training corpus, WorldArena 1.0, the four simulation benchmarks RoboTwin 2.0, CALVIN, RoboDojo, LIBERO and LIBERO-plus, and one packing task on the Green humanoid; the real-robot evaluation covers only this humanoid and this task.

The policies remain sensitive to changes in task geometry: a changed table height on RoboTwin and a changed robot initial configuration on LIBERO-plus both lower success, and the appearance augmentations used change observations without adding demonstrations of new motions. Long rollouts continue from the model's own generated frames, so prediction error accumulates from block to block; each benchmark is reached through its own post-training, and the real-robot evidence comes from one humanoid and one task.