Event

AWLS: An Offline Post-Training Method for Frozen One-Step Robot Manipulation Policies

1 reports · 1 sources

Jiahong Tang et al. propose Advantage-Weighted Latent Steering (AWLS), an offline post-training method that improves frozen one-step generative robot manipulation policies without updating the decoder; the paper was accepted to CoRL 2026. AWLS learns a residual noise shaper, i.e., an observation-conditioned residual correction to the noise in the policy's latent space, and trains it with an advantage-weighted action reconstruction objective; advantage values are clipped and scale-normalized and serve only as supervision weights for demonstration actions, avoiding policy updates via critic gradients while retaining fast single-step inference. The authors report that in experiments on simulated and real robot manipulation tasks, using clean and mixed-quality demonstrations, AWLS improves over the frozen base policy and limits drift from the pretrained behavioral prior.