DualVLA: Building a Generalizable Embodied Agent via Partial Decoupling of Reasoning and Action

TL;DR

DualVLA introduces a novel approach to improve generalizable embodied vision-language-action models by decoupling reasoning and action, employing data pruning and adaptive distillation to enhance performance without sacrificing reasoning capabilities.

Contribution

The paper proposes DualVLA, a method that enhances action performance in VLA models through data pruning and adaptive distillation, while maintaining reasoning ability, addressing the action degeneration problem.

Findings

Achieves 61.0 success rate in SimplerEnv

Scores 65.4 on average across eight benchmarks

Balances precise action execution with multimodal understanding

Abstract

To build a generalizable Vision-Language-Action (VLA) model with strong reasoning ability, a common strategy is to first train a specialist VLA on robot demonstrations to acquire reliable manipulation skills, and then incorporate mixed annotated robot data together with multimodal data to restore broader reasoning capabilities. However, we observe that the resulting reasoning VLA often suffers from degraded action performance compared to the specialist model before fine-tuning, a phenomenon we refer to as action degeneration. To address this issue, we propose DualVLA, which enhances action performance through carefully designed post-training while still preserving reasoning capability. We first introduce a dual-layer data pruning method that removes redundant embodied reasoning, preventing it from adversely influencing action learning. To further strengthen action generation, we design a dual-teacher adaptive distillation strategy that assigns different supervision signals to different data domains while maintaining reasoning ability. To fill the evaluation gap for generalist VLAs, we also propose VLA Score, which decouples VLA capability into reasoning, intention, action, and alignment dimensions for a more fine-grained assessment. Experiments show that DualVLA achieves an average success rate of 61.0 in SimplerEnv and an average score of 65.4 across eight competitive multimodal benchmarks, demonstrating a stronger balance between precise action execution and multimodal understanding. Project Website: https://costaliya.github.io/DualVLA/.