DREAMSTEER: Latent World Models Can Steer VLA Policies During Deployment Without Any Finetuning

2026-07-02

AI Summary

This work addresses the vulnerability of pretrained vision–language–action (VLA) policies to distribution shifts during deployment, which often leads to degraded robustness and inconsistent instruction following. The authors propose a novel, tuning-free, deployment-time guidance framework that dynamically steers VLA policies without modifying model parameters. By integrating an action-conditioned latent world model with a language-conditioned value model, the method enables zero-shot adaptation through candidate action-chunk sampling and trajectory ranking. Evaluated on four real-world manipulation benchmarks, the approach substantially improves task success rates from 23.75% to 66.25% and boosts instruction-following accuracy from 38.75% to 56.25%, demonstrating its effectiveness in enhancing policy generalization and alignment with high-level commands under distributional shift.

Abstract

Pretrained vision-language-action (VLA) policies show promising zero-shot generalization, but often fail under deployment-time distribution shift, leading to decreased robustness and inconsistent instruction following. While prior work commonly tackles this by finetuning on in-distribution data, it assumes demonstrations collected on tasks in the target environment. In this work, we propose DREAMSTEER, a deployment-time steering framework for pretrained VLAs without any finetuning or parameter modifications. The key insight in DREAMSTEER is to leverage a latent world model and a value model to steer pretrained VLA policies. During deployment, DREAMSTEER samples candidate action chunks from a VLA policy and predefined motion primitives, imagines their outcomes using an action-conditioned latent world model, and ranks the imagined trajectories with a language-conditioned value model. Across four real-world manipulation benchmarks with unseen objects, DREAMSTEER improves task success rate from 23.75% to 66.25% and instruction-following accuracy from 38.75% to 56.25% over the base VLA policy.

Problem

• distribution shift
• VLA policies
• instruction following
• deployment-time robustness

Innovation

• latent world model
• vision-language-action policy
• deployment-time steering
• zero-shot generalization
• value-guided planning