On a Unitree G1 humanoid, Long-WAM placed a moving green cup inside a blue cup in 19 of 20 trials. Two comparison controllers, π0.5 and Fast-WAM, succeeded in none of the same 20 attempts, the researchers report.
The task ran with the conveyor moving at 3 centimeters per second. A still image shows where each cup is; recent video helps the robot estimate how they are moving. Processing longer histories can help with motion, but it can also delay the next action.
Long-WAM first predicts what the video will look like next, then uses that prediction and the observed history to generate a chunk of robot movements. Its video model was trained to predict each next segment from earlier video, a method called autoregressive prediction. The team then adapted that model for robot actions while preserving the past-to-future structure.
The predicted future is stored as compressed visual representations, not rendered frames. That lets the action model use both what the robot has seen and what it expects to happen. To keep the longer video context from stalling control, the system encodes new observations while the robot executes its current movement chunk, then computes the next chunk asynchronously.
In a separate moving-cup grasping test, Long-WAM succeeded in 90% to 100% of trials as belt speed rose from 3.0 to 7.5 centimeters per second; each policy and speed condition used 20 trials.
On a separate RTX 5090 benchmark, the optimized system took 107.4 milliseconds per action chunk, including observation encoding and future-video prediction. The timing excludes preprocessing, text encoding and controller or communication overhead, and was not measured on the G1 robot itself.
The work is a new arXiv preprint. Physical comparisons used 20 trials per condition, so the result describes these controlled tasks rather than broad household reliability.
Sources
- Long-WAM paper and experiments. Figure 7 is used under the paper's CC BY 4.0 license.
- Research team project page and demonstration details.