Modular Sensory Stream for Integrating Physical Feedback in Vision-Language-Action Models

Jimin Lee1, Huiwon Jang1,2, Myungkyu Koo1,2, Jungwoo Park3, Jinwoo Shin1,2
1KAIST, 2RLWRLD, 3Seoul National University

TL;DR: MoSS enables Vision-Language-Action models to jointly leverage multiple physical feedback via modular sensory streams, yielding synergistic gains on contact-rich manipulation.

MoSS overview

Abstract

Humans understand and interact with the real world by relying on diverse physical feedback beyond visual perception. Motivated by this, recent approaches attempt to incorporate physical sensory signals into Vision-Language-Action models (VLAs). However, they typically focus on a single type of physical signal, failing to capture the heterogeneous and complementary nature of real-world interactions. In this paper, we propose MoSS, a modular sensory stream framework that adapts VLAs to leverage multiple sensory signals for action prediction. Specifically, we introduce decoupled modality streams that integrate heterogeneous physical signals into the action stream via joint cross-modal self-attention. To enable stable incorporation of new modalities, we adopt a two-stage training scheme that freezes pretrained VLA parameters in the early stage. Furthermore, to better capture contact interaction dynamics, we incorporate an auxiliary task that predicts future physical signals. Through extensive real-world experiments, we demonstrate that MoSS successfully augments VLAs to leverage diverse physical signals (i.e., tactile and torque), integrating multiple signals to achieve synergistic performance gains.

Method

MoSS is a modular adaptation framework for VLAs that enables scalable integration of multiple physical sensing modalities for contact-rich manipulation. The framework has two core components: (a) modular sensory streams, which process each physical signal (e.g., tactile, force, or torque) in parallel and inject them into the pretrained action expert via joint cross-modal self-attention, and (b) a two-stage training scheme, which first aligns the sensory streams with the frozen pretrained VLA and then jointly fine-tunes all components for coordinated control. In addition, an auxiliary future physical-signal prediction objective encourages the model to internalize contact dynamics, allowing heterogeneous sensory inputs to be combined in a synergistic and scalable manner.

MoSS method overview

Results

We present real-robot demonstrations for each task. Use the dropdown menu to select different task variants.

Task A Unstack Cup
Instruction "Pick up the top red cup from the stack and place it next to the blue cup."

GR00T N1.5 (Failure)

GR00T N1.5 + MoSS (Ours; Success)

Task B PnP Egg
Instruction "Pick up the egg and place it in the gold bowl."

GR00T N1.5 (Failure)

GR00T N1.5 + MoSS (Ours; Success)

Task C Board Erase
Instruction "Use the (red/grey) eraser to clean the whiteboard."

GR00T N1.5 (Failure)

GR00T N1.5 + MoSS (Ours; Success)

Task D Plug Insertion
Instruction "Pick up the (yellow/white/black) charger and plug it into the socket."

GR00T N1.5 (Failure)

GR00T N1.5 + MoSS (Ours; Success)

BibTeX

@misc{lee2026moss,
      title={Modular Sensory Stream for Integrating Physical Feedback in Vision-Language-Action Models},
      author={Jimin Lee and Huiwon Jang and Myungkyu Koo and Jungwoo Park and Jinwoo Shin},
      year={2026},
      eprint={2604.23272},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2604.23272},
}