VGM: Value-Gated Modulation for Minute-Level Residual Adaptation of VLA Policies

Zihan Wang, Yizhe Li, Zhe Han, Guoping Pan, Yi Cheng, Houde Liu

Keywords: VLA, residual adaptation, multi-modal perception, human-in-the-loop, rapid robot adaptation

TL;DR: We present VGM, a multimodal value-gated residual adaptation method that resolves intent misalignment and stagnation in frozen VLA policies, delivering rapid minute-level real-world adaptation and perfect task success with minimal demonstrations.

Abstract:

Vision-Language-Action (VLA) policies demonstrate remarkable semantic understanding, yet their real-world deployment is often hindered by execution failures in last-mile tasks. Residual learning offers a non-intrusive adaptation scheme to enhance performance without altering the base architecture; however, its effectiveness is often limited by a lack of coordination between the base model and the corrective branch. Addressing common challenges such as Intent Misalignment and Stagnation Traps in VLA adaptation, we propose the Value-Gated Modulator (VGM)—a framework designed for efficient, minute-level adaptation of generalist VLA policies.

The VGM leverages a multi-modal representation—integrating depth information alongside RGB and proprioception—to provide the necessary geometric grounding for precise interaction. Instead of unconstrained optimization, VGM functions as an adaptive arbitration layer that dynamically modulates the residual contribution by balancing strictly greedy optimization, preemptive regression detection, and forced recovery. Real-world experiments on a Franka robot arm demonstrate that VGM, initialized from only 20 demonstrations, enables a frozen VLA to converge to a 100\% success rate within under 5 minutes of real-world interaction—a $3.5\times$ speedup over static residual baselines. Our approach provides a practical and data-efficient pathway for deploying general-purpose robot policies in high-precision scenarios.

Submission Number: 11