On August 31, Triverse AI's joint research team publicly released a new embodied-AI result: Streaming-WAM. The official project credits researchers from Shanghai Jiao Tong University and Li Auto. The work aligns with Triverse AI’s long-term direction in world models and embodied intelligence and represents the latest public progress from our joint research ecosystem in real-time robot control.
Conventional World-Action Models (WAMs) often make the robot wait for the next visual and action sequence before it can continue. Streaming-WAM turns this synchronous loop into streaming control: while the robot executes the current action chunk, the model prepares the next video and action predictions in the background, reducing stalls caused by inference.
Keeping prediction aligned with motion already underway
Naive asynchronous execution creates a timing mismatch: the robot keeps moving after inference starts, so the imagined future may fall behind the real state. Streaming-WAM conditions its world model on the committed action prefix, allowing future visual prediction to explicitly account for the motion already underway and produce smoother, better-aligned transitions between chunks.
Key results
- LIBERO: 98.20% average success. Compared with FastWAM, chunk time falls from 493.0 ms to 41.0 ms, a 12.0× speedup, while long-task episode time drops from 16.31 s to 5.36 s.
- RoboTwin 2.0: overall success rises from 87.00% with FastWAM-Joint to 91.74%, while chunk time falls from 652.1 ms to 54.4 ms.
- Real robot: 27 successes in 30 trials, or 90.00%; one rollout falls from 90 s to 38 s, a 2.4× end-to-end speedup.
Open-source status
Training and inference code, the accelerated runtime, LIBERO / RoboTwin recipes, and checkpoints are available under the Apache-2.0 License. The technical report is forthcoming.