文章背景与核心概要

视觉-语言-动作（VLA）模型在将多模态输入转化为机器人动作方面表现出色，但其动作解码器主要仍通过标准的行为克隆进行训练。这种方法仅监督低-level（底层）的运动指令，却将行为局部的语义目标隐式化，导致模型难以捕捉复杂任务中长期的行为意图。

为了解决这一局限性，本文作者提出了意图蒸馏（Intention Distillation, INDI）方法。在训练过程中，冻结的教师视觉-语言模型（VLM）结合当前的观测、指令、粗粒度动作摘要以及执行视频来解释所演示的片段。部署后的 VLA 随后在中间解码器层恢复这一多模态意图表示，从而引导动作预测。实验表明，INDI 在 SimplerEnv-Bridge、RoboCasa 厨房环境以及真实世界任务中均实现了显著的性能提升。

Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

Authors: Sangoh Lee, Sangwoo Mo, Wook-Shin Han
Subjects: Robotics (cs.RO); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
arXiv: 2608.23478 | Submitted: August 24, 2026
Project Page: Link

📌 Summary

Vision-Language-Action (VLA) models excel at translating multimodal inputs into robotic actions, but their action decoders are predominantly trained using standard behavior cloning. This approach supervises low-level motor commands while keeping the local semantic objective of the behavior implicit.

To address this limitation, the authors propose Intention Distillation (INDI). During training, a frozen teacher VLM interprets demonstrated segments using current observations, instructions, coarse action summaries, and execution videos. The deployed VLA then recovers this multimodal intent representation at an intermediate decoder layer to guide action prediction.

Key Results: * SimplerEnv-Bridge: Improves GR00T-N1.7 performance from 64.3% to 84.7%. * RoboCasa Kitchen: Boosts the baseline from 64.1% to 70.3%. * \(\pi_{0.5}\) Benchmarks: Shows consistent performance gains across multiple benchmarks. * Real-World Tasks: Raises average task success rate from 62.0% to 68.7%, with up to a 12.0 percentage point (pp) improvement on longer-horizon tasks.

🔬 Abstract

Vision-Language-Action (VLA) models can turn multimodal context into robot actions, but their action decoders are still trained largely by behavior cloning. This supervises which motor command was demonstrated while leaving implicit the local objective served by the behavior under the instruction. Future-based supervision enriches action learning with frames, latent observations, trajectories, or motion representations, but these signals capture particular realizations of what may happen rather than the shared semantic objective of the forthcoming behavior.

We propose Intention Distillation (INDI), which distills behavior-level intent into the action decoder. During training, a frozen teacher VLM interprets a demonstrated segment from the current observation, instruction, coarse action summary, and corresponding execution video. From its standard inputs, the deployed VLA recovers the resulting multimodal intent representation at an intermediate decoder layer and uses it to organize action prediction together with representations of how the behavior unfolds and what it achieves.

These results show that action decoders benefit from explicitly modeling the semantic objective of the behavior they generate.