Abstract
Vision-language-action (VLA) models offer a promising paradigm for surgical robot autonomy by mapping visual observations and language instructions to continuous control. However, existing surgical VLA systems remain limited in manipulation complexity and often suffer from instability and compounding errors in long-horizon settings, particularly
when trained solely on demonstrations. Most are optimized for single procedural objectives rather than unified control across diverse procedures. To address this gap, we present SurgVLA-Struct, a reinforcement refined framework for surgical robotic manipulation across tasks in the SurRoL simulator. Our approach comprises three key components: (1) a structured 10-task benchmark with 7 newly introduced language-conditioned variants incorporating obstacle constraints, extended horizons, and attribute
driven instructions; (2) a unified continuous-control backbone with chunked 7-DoF action prediction under a shared action space, enabling stable and consistent language-conditioned manipulation without task-specific decoders; and (3) structured surgical reward modeling within a GRPO framework using task-invariant, simulator-verifiable objectives encoding precision, efficiency, trajectory stability, and contact safety. Extensive experiments show that SurgVLA-Struct consistently outperforms classical reinforcement learning baselines and state-of-the-art VLA models, validating stable language-conditioned manipulation across diverse tasks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COLAS_002.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=cgJXCBnzM8
BibTex
@InProceedings{ZhaErl_SurgVLAStruct_MICCAISAT2026,
author = { Zhang, Erli AND Yang, Shuojue AND Shen, Daiyun AND Li, Zhengyang AND Jin, Yueming},
title = { { SurgVLA-Struct: Structured Surgical Reward Modeling for Across-Task Vision-Language Robotic Manipulation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS pending},
month = {pending},
page = {pending}
}