Unified Noise Steering for Efficient Human-Guided VLA Adaptation

2026-05-11

AI Summary

Efficient fine-tuning of pretrained vision-language-action (VLA) models on real robots remains hindered by costly online interactions and inefficient exploration. This work proposes UniSteer, a framework that unifies human action interventions with noise-space reinforcement learning for the first time. By leveraging a frozen flow-matching decoder to approximate the inverse mapping from actions to noise, human corrections are transformed into supervision signals in the noise space, enabling joint optimization of a lightweight noise predictor. Evaluated on four real-world manipulation tasks, UniSteer achieves an average success rate increase from 20% to 90% with only 66 minutes of human-robot interaction, substantially outperforming existing baselines in both noise-space reinforcement learning and action-space human-in-the-loop collaboration.

Abstract

Diffusion-based vision-language-action (VLA) models have emerged as strong priors for robotic manipulation, yet adapting them to real-world distributions remains challenging. In particular, on-robot reinforcement learning (RL) is expensive and time-consuming, so effective adaptation depends on efficient policy improvement within a limited budget of real-world interactions. Noise-space RL lowers the cost by keeping the pretrained VLA fixed as a denoising generator while updating only a lightweight actor that predicts the noise. However, its performance is still limited due to inefficient autonomous exploration. Human corrective interventions can reduce this exploration burden, but they are naturally provided in action space, whereas noise-space finetuning requires supervision over noise variables. To address these challenges, we propose UniSteer, a Unified Noise Steering framework that combines human corrective guidance with noise-space RL through approximate action-to-noise inversion. Given a human corrective action, UniSteer inverts the frozen flow-matching decoder to recover a noise target, which provides supervised guidance for the same noise actor that is simultaneously optimized via reinforcement learning. Real-world experiments on diverse manipulation tasks show that UniSteer adapts more efficiently than strong noise-space RL and action-space human-in-the-loop baselines, improving the success rate from 20% to 90% in 66 minutes on average across four real-world adaptation tasks.

Problem

• vision-language-action
• noise-space RL
• human-in-the-loop
• robotic manipulation
• policy adaptation

Innovation

• noise-space RL
• human-in-the-loop
• action-to-noise inversion
• diffusion-based VLA
• efficient adaptation