Stanford Robotics Seminar ENGR319 | Spring 2026 | Unlocking Autonomous Medical Robotics

Current surgical robots are predominantly teleoperated, meaning they lack inherent skill and rely entirely on the surgeon. While they enhance precision, they do not address the critical shortage of medical personnel or allow for performing more surgeries.

Surgical robotics presents challenges like scarce data, uncontrolled and non-resettable environments, extremely high safety requirements, few demonstrators, and privacy concerns around patient data, which are absent in cleaner, low-stakes domains where current foundation models have been developed.

The four foundational pillars are perception (seeing and feeling the world), modeling and simulation (understanding physics and how the world evolves), planning (making decisions based on world models), and control (executing actions precisely).

Surgical environments pose unique perception challenges including narrow field of view, poor depth measurements, specular reflections from fluids, smoke, constant occlusion by instruments, deformable anatomy, difficulty finding static anchor points, and insufficient data for training modern neural models.

Techniques like deformable reconstruction from multiple cameras and scene fusion allow robots to build a 3D understanding of the surgical site, even as anatomy is manipulated. This is crucial for tasks like tumor excision with margin awareness and for understanding physics.

Digital twinning involves creating a physics-based simulation of deformable tissue that can run in real-time. This allows robots to simulate potential interactions and consequences before acting, informing model-based control and improving accuracy through techniques like differentiable rendering.

It's an approach that combines engineered or learned behaviors into 'knowledge modules' within a neural network. Through reinforcement learning, the system learns how to sequence and combine these modules to achieve longer, more complex tasks, enabling lifelong learning.

Humanoid robots offer a versatile platform that can potentially perform a wide range of tasks currently handled by multiple specialized robots or human assistants, such as remote surgery, autonomous assistance, and nursing tasks, potentially at a lower cost than multiple dedicated systems.

Teleoperating robot hands is challenging due to limited degrees of freedom, size mismatches, and critically, limited tactile sensing. Improvements involve using reinforcement learning to teach robots to grasp instruments autonomously and developing advanced haptic feedback systems.

A significant bottleneck is capturing robust force and tactile data. This is hampered by limited interfaces for sensing and material science limitations in tactile sensors (sensitivity vs. robustness trade-off), making it a long-term challenge.

Robots can incorporate metrics related to pulling force and tissue stretching to define safe boundary conditions. By integrating this into control systems, they can predict and avoid excessive strain that might lead to tearing, informed by both simulation and visual feedback.