Safe Uncertainty-Aware Learning of Robotic Suturing

TL;DR

Safe uncertainty-aware autonomous robotic suturing is addressed by combining an ensemble of diffusion policies with Out-of-Distribution detection to decide when to hand control to a surgeon. The approach adds a model-free Control Barrier Function to impose formal safety constraints on the robot controller. The paper demonstrates that the diffusion-policy ensemble provides meaningful epistemic uncertainty estimates and that the CBF can constrain actions to a predefined safe set even under unsafe predictions. Evaluations in a surgical simulator show robust task execution under perturbations such as needle drops, camera shifts, and phantom movements, with OOD detection enabling timely human intervention.

Abstract

Robot-Assisted Minimally Invasive Surgery is currently fully manually controlled by a trained surgeon. Automating this has great potential for alleviating issues, e.g., physical strain, highly repetitive tasks, and shortages of trained surgeons. For these reasons, recent works have utilized Artificial Intelligence methods, which show promising adaptability. Despite these advances, there is skepticism of these methods because they lack explainability and robust safety guarantees. This paper presents a framework for a safe, uncertainty-aware learning method. We train an Ensemble Model of Diffusion Policies using expert demonstrations of needle insertion. Using an Ensemble model, we can quantify the policy's epistemic uncertainty, which is used to determine Out-Of-Distribution scenarios. This allows the system to release control back to the surgeon in the event of an unsafe scenario. Additionally, we implement a model-free Control Barrier Function to place formal safety guarantees on the predicted action. We experimentally evaluate our proposed framework using a state-of-the-art robotic suturing simulator. We evaluate multiple scenarios, such as dropping the needle, moving the camera, and moving the phantom. The learned policy is robust to these perturbations, showing corrective behaviors and generalization, and it is possible to detect Out-Of-Distribution scenarios. We further demonstrate that the Control Barrier Function successfully limits the action to remain within our specified safety set in the case of unsafe predictions.

Figures (12)

• Figure 1: Our proposed framework consists of 1) a Diffusion Policy Ensemble model, which utilizes multiple trained Diffusion Policy models for quantifying epistemic uncertainty in the learned policy, 2) an O.O.D. detection scheme that is used to release control back to the surgeon, depeding on the uncertainty in the policy. It is possible to detect O.O.D. uncertainties based on a calibration test using this uncertainty measure, and 3) A Control Barrier Function, which acts as a safety filter on the robot controller and adds redundancy to safety measures by enforcing manually placed safety boundaries.
• Figure 2: Illustrations of how the diffusion process is applied on the demonstration trajectories and how the model trains to reverse the diffusion process
• Figure 3: Image sequence showing a needle insertion demonstration.
• Figure 4: Image sequence of a single suture throw execution with a model ensemble consisting of four trained diffusion policy models' predictions visualized.
• Figure 5: The top graph depicts the estimated uncertainty ($\max (\Sigma)$) for each prediction at each timestep by the model ensemble. The background indicates the prediction of the O.O.D. detector, i.e., green for I.D. and red for O.O.D.. We execute the scenario 12 times. We use the maximum sample of the 12 runs during the O.O.D detection. The bottom images show that the scenario starts with I.D. needle insertion; however, at timestep 8, the needle is dropped and detected at timestep 30.

Theorems & Definitions (2)

• Theorem 1: molnar2021model
• Corollary 1: molnar2021model