Event
GPL Represents VLA Inference State as a Plan Distribution
1 reports · 1 sources
Matteo Cercola proposes Generative Plan Latents (GPL), which represent the inference state of vision-language-action (VLA) models as a low-dimensional conditional distribution over plans (e.g., which object to use, grasping, or route). The authors claim that this allows plans to be sampled (coverage), selected on request (guidance), and reweighted according to deployment outcomes (outcome adaptation) without retraining the VLA. Experiments test these three capabilities with π₀ on SimplerEnv manipulation tasks and extend to RoboCasa, LIBERO, and a maze domain. The authors report that guided plans raise the proportion of executions that follow the request from 0.48 to 0.94, whereas applying the same guidance to the native action distribution shows no reliable gain.
Report timeline
- GPL: Representing VLA reasoning state as a plan distribution
The authors propose Generative Plan Latents (GPL), which represent the reasoning state of vision-language-action (VLA) models as a low-dimensional conditional distribution over plans, such as which object, grasp, or route to use, so that plans can be sampled, selected on request, and reweighted by deployment outcomes without retraining the VLA.
arXiv cs.RO ·