Tech • AI • Robotics
GTC SJ 2026: Physical AI for Healthcare Robotics - Simulation-First Design & Accelerated Development
7/10TL;DR
Healthcare robotics is moving from concept to deployment as hospitals, researchers and device makers use foundation models, synthetic data and higher-fidelity simulation to tackle surgical autonomy, training and operating-room efficiency.
KEY POINTS
A supply-and-demand crisis is driving adoption
Healthcare systems face rising patient demand and limited clinical capacity, making robotics a potential lever for speed, consistency and accuracy. The challenge is especially acute because healthcare environments are highly dynamic, from changing anatomy during surgery to constantly shifting hospital workflows, which requires more generalizable AI models than many industrial settings.
Data remains the central bottleneck
Progress depends on a full data pipeline: curation, annotation, search, gap analysis, synthetic generation, training and deployment. In practice, each step is labor intensive, and even similar clinical departments can label the same data differently. Rare but critical moments such as bleeding or accidental injury may account for only 2% to 3% of a surgical video, making it difficult to gather enough examples of the highest-risk events.
Three AI shifts are reshaping healthcare robotics
The sector is drawing on three broader advances in physical AI: foundation models that create physically plausible data, reasoning models that can explain actions, and faster, more realistic simulation. Developers increasingly start from open models, add proprietary hospital or device data, and then supplement it with either synthetic generation or first-principles simulation before post-training and testing.
Robotic surgery has created a new data source
Surgeons described the robot as the biggest change in the operating room since laparoscopy. Robotic systems now hold more than 30% of the minimally invasive surgery market, while also capturing rich streams of video and kinematic data. Much of that information is still underused, but researchers are beginning to turn it into performance models aimed at improving surgeon training and, eventually, enabling autonomous functions.
Surgical annotation is becoming more structured
Hospitals and research groups are building shared surgical ontologies to classify anatomy, procedural phases and gestures. Those gesture labels are especially important for video-language-action models because they connect what the camera sees to what the robot should do. Video-language models are also reducing the time needed to annotate large surgical datasets that previously depended on scarce clinician reviewers.
Simulation is being redesigned for surgery
Conventional surgical simulators render scenes pixel by pixel, but researchers are testing approaches that begin with a real surgical frame and then simulate tissue interaction using robot kinematics and learned models. The goal is to produce more realistic, scalable training environments for surgeons and robots alike. Advocates argue this is not just a technical upgrade but an ethical need, because many trainees still learn complex robotic skills on live patients rather than in sufficiently capable simulation.
Autonomous surgery is advancing through task-level milestones
A Johns Hopkins team reported imitation-learning systems that can perform core robotic tasks such as tissue lifting, needle pickup, handover and knot tying, some with 100% success rates in controlled studies. The same approach has been extended to a more complex phase of gallbladder removal, where an autonomous system clipped and cut the bile duct and cystic artery in an ex vivo porcine model. Researchers said earlier model-based systems plateaued near a 60% stitch-to-stitch success rate, while learning-based methods improved robustness and scalability.
A major open dataset is accelerating development
Researchers and partners released what they described as the largest published surgical robotics dataset, with more than 150,000 trajectories and about 1 terabyte of data. The collection had already drawn roughly 1,000 downloads shortly after release. Early fine-tuning results on GROOT-based models suggested stronger suturing performance than prior systems and high proficiency using only 33% of the suturing dataset.
Commercial systems are using AI first on lower-risk workflow problems
Moon Surgical said its Maestro platform, a two-arm cart-based laparoscopic robot available in the US and EU, has been used in nearly 3,000 procedures. Rather than starting with high-risk autonomy, the company is focusing on setup optimization, arm deployment and camera following, including its Scope Pilot feature that lets a surgeon effectively control three instruments with two hands. Simulation in Isaac and synthetic scene generation with Cosmos Transfer are being used to model room layouts, patient draping, bed positions and robot placement.
Orthopedic robotics is pursuing a different architecture
Lamb Surgical, a Swiss company with FDA clearance, argued that hard-tissue robotics has lagged soft-tissue surgery because most systems still rely on a single robotic arm and navigation camera. Its answer is an upper-torso, three-arm architecture with two operating arms plus synchronized vision, designed to work with nonproprietary tools and eventually support supervised autonomy. The company said the urgency is growing because 40% of US orthopedic surgeons are already older than 60, raising concerns about future capacity as the elderly population expands.
CONCLUSION
Healthcare robotics is shifting from isolated prototypes to integrated AI pipelines built on real, synthetic and simulated data. The main contest now is not whether autonomy will enter care, but how quickly developers can make it safe, explainable and scalable in the operating room and beyond.
Explain this