MoE-ACT: Improving Surgical Imitation Learning Policies through Supervised Mixture-of-Experts

TL;DR

This work tackles data-efficient learning for autonomous surgical manipulation by coupling Action Chunking Transformer (ACT) with a supervised Mixture-of-Experts (MoE) architecture that exploits phase structure. The approach uses a phase-aware MoE block atop a lightweight, vision-only policy trained with a 4-term objective, enabling robust long-horizon tasks from fewer than 150 demonstrations using stereo endoscopic imagery. Empirical results on a bowel grasping and retraction task show that MoE-ACT outperforms baseline ACT and state-of-the-art Vision-Language-Action models, with strong in-distribution performance, improved out-of-distribution robustness, and a successful 80% zero-shot transfer to ex vivo porcine tissue, plus preliminary in vivo results. The findings highlight the value of explicit phase supervision and viewpoint variability for safe, real-time surgical automation and point toward practical in vivo deployment as a future direction.

Abstract

Imitation learning has achieved remarkable success in robotic manipulation, yet its application to surgical robotics remains challenging due to data scarcity, constrained workspaces, and the need for an exceptional level of safety and predictability. We present a supervised Mixture-of-Experts (MoE) architecture designed for phase-structured surgical manipulation tasks, which can be added on top of any autonomous policy. Unlike prior surgical robot learning approaches that rely on multi-camera setups or thousands of demonstrations, we show that a lightweight action decoder policy like Action Chunking Transformer (ACT) can learn complex, long-horizon manipulation from less than 150 demonstrations using solely stereo endoscopic images, when equipped with our architecture. We evaluate our approach on the collaborative surgical task of bowel grasping and retraction, where a robot assistant interprets visual cues from a human surgeon, executes targeted grasping on deformable tissue, and performs sustained retraction. We benchmark our method against state-of-the-art Vision-Language-Action (VLA) models and the standard ACT baseline. Our results show that generalist VLAs fail to acquire the task entirely, even under standard in-distribution conditions. Furthermore, while standard ACT achieves moderate success in-distribution, adopting a supervised MoE architecture significantly boosts its performance, yielding higher success rates in-distribution and demonstrating superior robustness in out-of-distribution scenarios, including novel grasp locations, reduced illumination, and partial occlusions. Notably, it generalizes to unseen testing viewpoints and also transfers zero-shot to ex vivo porcine tissue without additional training, offering a promising pathway toward in vivo deployment. To support this, we present qualitative preliminary results of policy roll-outs during in vivo porcine surgery.

Figures (10)

• Figure 1: Phantom, Ex Vivo and In Vivo Porcine Bowel Grasping and Retraction, policy roll-outs.
• Figure 2: Experimental setup using the OpenHELP open-body phantom, showing the phantom with two robotic arms, one holding the laparoscope and one holding the surgical instrument. The abdominal wall cover is removed for visibility purposes.
• Figure 3: Policy architecture: ACT is extended with a MoE block.
• Figure 4: Roll-outs of our policy trained on the random-viewpoint dataset generalize on unseen camera viewpoints, showing robust performance across zoom and orientation changes. Examples show initial (left) and final (right) frames of the roll-outs.
• Figure 5: Ablation on the amount of training data demonstrations plotted against the policy success rates for In-Distribution Roll-Outs