Event
ReStruct inference-time guidance for robot policies accepted to CoRL 2026
1 reports · 1 sources
ReStruct, proposed by Yiyuan Pan et al., has been accepted to CoRL 2026. The method guides learned robot policies at inference time, enabling policies to satisfy user preferences not specified during training without retraining: it decomposes visuomotor policies into a high-level state-machine skeleton and a low-level residual controller, and updates the action prior provided by the skeleton while keeping the controller frozen, thereby achieving physically aware control under altered task structures. According to the paper, simulation and real-world experiments show that ReStruct can guide a variety of preferences ranging from object-level specifications to temporal-logic constraints, and after guidance it surpasses existing methods in task success rate and preference adherence, leading VLA models by up to 25%.
Report timeline
- ReStruct: Guiding Robot Behavior at Inference Time via Physically-Aware Reconfiguration of Task Structure
ReStruct proposes a method to steer learned robot policies at inference time, enabling them to satisfy user preferences not specified during training without retraining. The method decomposes a visuomotor policy into a high-level state-machine skeleton and a low-level residual controller, and incorporates preferences into the skeleton via a synchronous product to reconfigure the task structure, while the controller remains frozen. Simulation and real-world experiments show that ReStruct can handle a variety of preferences, from object-level specifications to temporal-logic constraints, and after steering it outperforms existing methods in task success rate and preference adherence, leading VLA models by up to 25%.
arXiv cs.RO ·