Execution-Aligned Progressive Noise for Consistent Asynchronous Replanning in Generative Robot Policies

Execution-Aligned Progressive Noise for Consistent Asynchronous Replanning in Generative Robot Policies. By Di Wu and five others, from Magiclab Robotics and Southeast University. Published on the fifth of October. No code repository is linked to it.

A generative robot policy working in real time has to keep replanning while it acts, continuously and asynchronously, producing one chunk of actions after another. Each new chunk usually starts from fresh, independent random noise. That independence is the trouble. A chunk drawn from unrelated noise can switch to a different mode of behaviour from the one before it, and the robot's continuation from chunk to chunk becomes inconsistent.

Execution-Aligned Progressive Noise, or EAPN, replaces independent initialization with structured stochasticity, at two levels. Between chunks, the noise is shared and carried forward. Within a chunk, it is correlated along action time.

Across replanning steps, EAPN propagates a single shared noise trajectory and aligns it with the displacement the robot has actually executed, which gives consecutive chunks a correlation tied to real execution. Inside each chunk, it models temporal correlation along the action sequence. This aligned stochastic history is then combined with the context of actions already committed, and the two together condition what is generated next. A new chunk continues from a generative state consistent with execution, instead of restarting from independent noise.

The evaluation covers D3IL, Kinetix, LIBERO, and real-world manipulation. On D3IL, EAPN improves the consistency of multimodal behaviour. On Kinetix it reaches an average success rate of eighty-eight point five nine percent. On LIBERO it stays robust, maintaining strong task performance even under long inference delays. On a real robot, it succeeds ninety percent of the time on Object Storage, and ninety-six point seven percent of the time on bimanual Cloth Folding.

The authors take these results as showing reliable continuous execution under asynchronous replanning. The change is small to state. Randomness, which used to be drawn anew each time the robot replanned, now has a memory, and that memory is anchored to what the robot actually did.