Products

Streaming-WAM: Predicting Ahead While the Robot Keeps Moving

Streaming World-Action Model for Real-Time Robot Manipulation

Streaming-WAM turns world-model inference and robot execution into a parallel streaming system: while the robot executes its current action chunk, the model prepares the next visual and action predictions, substantially reducing inference-induced stalls.

Core Design

• ▸ Streaming generation: generates the next world-action chunk while the current action is executing
• ▸ Action-conditioned future prediction: conditions prediction on committed actions to align the imagined future with motion already underway
• ▸ Consistency distillation: compresses generation latency while action conditioning preserves task performance

Measured Results

Across LIBERO, RoboTwin 2.0, and a real robot, Streaming-WAM demonstrates low latency, continuous control, and strong task success. Code, runtime, recipes, and checkpoints are now open source.