Think Proprioceptively: Compact Subgoal Traces for Vision-Language-Action Model

Fangyuan Wang, Peng ZHOU, Shipeng Lyu, Gu Gong, Weiwei Lin, Anqing Duan, David Navarro-Alarcon, Guodong Guo

Keywords: planning, vla, robotics, subgoal, manipulation

Abstract: Vision-language-action (VLA) models translate visual observations and language instructions to robot actions, yet current architectures regard proprioception as a passive input rather than an active reasoning component. Without proprioceptive guidance, VLA models process multimodal features in isolation from the robot’s physical configuration, and hierarchical approaches often encode subgoals in high-dimensional visual or textual spaces that are ungrounded in the robot’s embodiment. We present SubgoalVLA, a framework built on the \textit{think proprioceptively} paradigm that redefines how multimodal information is processed. SubgoalVLA leverages proprioception in two ways. First, proprioceptive states serve as cross-attention queries to select vision-language features, enabling configuration-aware feature extraction. Second, subgoals are encoded as compact sequences of joint configurations that eliminate the need for cross-modal translation. Through a two-stage training protocol that begins with supervised learning on ground-truth subgoals and then fine-tunes with self-predicted subgoals, we mitigate distribution shift between training and inference. On the CALVIN benchmark, SubgoalVLA achieves state-of-the-art performance with an average task completion length of 3.32, demonstrating that proprioceptive reasoning provides the critical bridge between high-level task understanding and embodied control.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 3710