Failure Detection for Surgical Robot Imitation Policies via Flow-Matching World Modeling
This research paper introduces FoMo-FD (Flow-Matching World Model for Failure Detection), a method that learns nominal short-horizon visual dynamics with an action-conditioned flow-matching world model to detect low-level execution failures in surgical robot imitation policies. By scoring the inverse-transport nonconformity of observed endpoint latents and calibrating detection thresholds on successful rollouts, the method achieves robust window-level failure detection across multiple tasks without requiring failure demonstrations. Evaluations on simulation and real-world da Vinci Research Kit (dVRK) manipulation tasks demonstrate high failure detection rates with low false alarm rates.
Summary and content fields are drawn from the document text and are pending editorial review.
Who issued it
Who it governs
Researchers and developers working on autonomous robotic surgery, imitation learning, and runtime safety monitoring.
What the document contains
The document covers the formulation, architecture, and evaluation of an endpoint-prediction flow-matching world model for runtime failure detection in surgical robot imitation policies. It addresses window-level execution monitoring, inverse-transport nonconformity scoring, and conformal threshold calibration using multi-view observations across simulated and real-world manipulation tasks.
The text does not provide a formal regulatory framework, clinical deployment guidelines, or safety compliance certifications for commercial medical devices.
- FoMo-FD uses an action-conditioned flow-matching world model to predict endpoint latents of short execution windows in surgical robotic tasks.
- Failure detection is formulated as scoring the inverse-transport nonconformity of observed endpoint latents relative to learned nominal dynamics.
- Detection thresholds are calibrated using only successful policy rollouts via conformal prediction, eliminating the need for failure demonstrations.
- Experiments across simulation and real-world dVRK tasks show that the wrist-camera view achieves superior failure detection compared to fixed-camera views.
- Action conditioning and multi-step endpoint prediction horizons substantially improve detection performance while maintaining low false alarm rates.
· I. Introduction · II. Problem Formulation · III. Method · IV. Experiments · V. Ablation Study · VI. Discussion and Conclusion
How it scores
Each score is backed by a quote verified against the document text. Editorial review has not yet confirmed them.
Accountability
0/5Clear responsibility for AI outcomes
N/A
screening score · 2026-10-24
Fairness
0/5Non-discrimination and equitable treatment
N/A
screening score · 2026-10-24
Human Dignity
0/5Respect for the inherent worth of persons
N/A
screening score · 2026-10-24
Human Oversight
2/5Humans remain in control of consequential decisions
Autonomy in robot-assisted surgery has the potential to improve procedural consistency, reduce surgeon workload, and expand access to high-quality care.
Section I
screening score · 2026-10-24
Privacy
0/5Protection of personal data and private life
N/A
screening score · 2026-10-24
Transparency
4/5Openness about how AI systems work and are used
Source code, trained models, and processed datasets will be released upon acceptance.
Abstract
screening score · 2026-10-24
Enforcement
0/5Consequences and compliance mechanisms with real force
N/A
screening score · 2026-10-24
Impact Assessment
4/5Required assessment of an AI system's risks/impacts before or during use
Failure detection is therefore an essential safeguard, but its development remains difficult due to the challenges of scarce failure data, highly variable manipulation dynamics, and the need to balance missed detections against disruptive false alarms.
Abstract
screening score · 2026-10-24