Execution-Aligned Progressive Noise for Consistent Asynchronous Replanning in Generative Robot Policies
EAPN keeps robot action chunks coherent by carrying execution-aligned noise forward instead of restarting each replan independently.
The paper targets real-time generative robot policies where asynchronous replanning can create mode switches between action chunks. Its method adds structured stochasticity across replans and within each chunk, using actual execution displacement and committed action context to condition the next generation. The authors report gains on D3IL, 88.59% average success on Kinetix, robustness under long inference delays on LIBERO, and 90.0% / 96.7% success in two real-robot tasks. HF Daily Papers' note
The paper targets real-time generative robot policies where asynchronous replanning can create mode switches between action chunks. Its method adds structured stochasticity across replans and within each chunk, using actual execution displacement and committed action context to condition the next generation. The authors report gains on D3IL, 88.59% average success on Kinetix, robustness under long inference delays on LIBERO, and 90.0% / 96.7% success in two real-robot tasks. HF Daily Papers' note
score 4