CASE STUDY
How a structured teleoperation and annotation program helped a surgical robotics team cut tissue contact errors by 67% in clinical simulation benchmarks.
A surgical robotics company was preparing for clinical trials of an assistive surgical system. Their autonomous mode was performing below threshold on the key safety metric: unintended tissue contact rate. The model needed better demonstration data from expert-level surgical simulation scenarios.
Surgical robot training data requires operators with specialist knowledge, a clinical simulation environment, and a quality framework that meets the regulatory standards the end system will be evaluated against. The team had tried crowd-sourced annotation but the quality was insufficient for the safety-critical task. They needed a structured program with credentialled operators and rigorous QA.
Roborax deployed a dedicated team of trained surgical simulation operators working in a clinical simulation center. Over eight weeks, we captured 1,200 expert demonstration episodes across 12 surgical task types, with synchronized force-torque, video, and instrument state data. Each episode was reviewed by a QA lead with clinical simulation experience before delivery.
The client’s model, retrained on the Roborax dataset, achieved a 67% reduction in tissue contact errors on the clinical benchmark suite — bringing performance above the threshold required for the next phase of trials.
Related: Surgical and medical solutions — Case studies.
In safety-critical robotics applications, the quality of demonstration data directly determines the safety properties of the resulting policy. A model trained on mediocre demonstrations will learn mediocre behavior. A model trained on expert-level demonstrations — captured by operators who understand the task, reviewed by QA leads who understand the safety requirements — has the foundation to achieve expert-level performance. Related: Surgical and medical solutions — Case studies.
