Where physics meets machine learning. We build world models that give surgical robots the understanding of tissue, force, and sound -- so every motion is safe by construction.
Michelangelo painted the moment before contact -- fingertips reaching across the void. That gap holds everything: intention, potential, the instant before creation begins.
We named ourselves for that space. Hands Robotics exists in the gap between the physical world and the intelligence to understand it. Our graph neural networks don't approximate physics. They learn it -- from the topology of contact, the propagation of force, the resonance of material.
Real-time soft tissue simulation with safety by construction. Encodes objects as particle graphs and propagates contact forces through message-passing layers. Validated on 7 tissue types -- brain, liver, kidney, myocardium, vessel, cartilage, bone. Zero training data needed.
Differentiable wave equation implemented as GNN message passing. Synthesizes CT-to-ultrasound images without paired data. Sub-millimeter accuracy for intraoperative guidance. The machine learns to hear what it touches.
Combines tissue deformation prediction with ultrasound guidance in a single pipeline. The surgeon sees beneath the surface -- organ shift tracked in real time, instrument paths verified against the world model before execution.
The unifying layer. Couples mechanical, acoustic, and sensing physics into a single differentiable graph. One model that understands how the world behaves across domains -- differentiable end-to-end for robot training.
Traditional world models learn physics from data — and inherit every bias, gap, and failure mode of that data. We took a fundamentally different path. In DPC-GNN, Newtonian mechanics, conservation laws, and material constitutive equations are not loss terms or soft regularizers — they are the architecture itself. Every message-passing layer enforces momentum conservation through antisymmetric force computation. Every node update respects energy bounds by construction. The result: a world model that requires zero training data yet achieves sub-millimeter accuracy across 7 biological tissue types — brain, liver, kidney, myocardium, vessel wall, cartilage, and bone. No domain gap. No hallucinated dynamics. No catastrophic failures when encountering unseen tissue geometries. For surgical robotics, this is not a technical preference — it is a safety imperative. When a robot touches living tissue, the physics must be right by construction, not by probability.
We encoded the acoustic wave equation directly as differentiable message-passing operations — a world first. Where existing methods approximate ultrasound with learned CNNs or soft-constraint loss functions, our approach hard-wires the Helmholtz equation into the GNN layer computation. Sound propagation through heterogeneous tissue is physically guaranteed, not statistically estimated. This enables real-time CT-to-ultrasound synthesis without any paired CT-US training data — the model generates physically correct ultrasound purely from tissue properties and geometry. For surgeons, this means seeing beneath the surface with physics-guaranteed accuracy. For the field, this proves that acoustic simulation belongs in the same graph framework as mechanical simulation — a unification that opens entirely new possibilities for multi-modal surgical perception.
The gradient dead zone — where soft tissue dynamics produce vanishing or exploding gradients — is the fundamental reason reinforcement learning fails in surgical robotics. Standard RL treats the environment as a black box and wastes millions of samples learning what physics already knows. Our DPC-GNN world model provides analytically differentiable gradients through the entire physics simulation, eliminating the gradient dead zone entirely. DiffPPO — our physics-accelerated policy optimization — converges orders of magnitude faster than standard PPO because every gradient respects conservation laws and material response. The Multi-Physics World Model (MPWM) unifies mechanics, acoustics, and thermodynamics in a single differentiable graph, enabling robot training across all physical modalities simultaneously. This is not just faster training. This is the only path to surgical autonomy where safety is mathematically guaranteed at every optimization step.
FDA cleared. First-in-human cancer treatment completed. Our Cold Atmospheric Plasma (CAP) device delivers precise, controlled energy to target tissue -- enabling selective cancer cell destruction while preserving healthy tissue.
Modular architecture: plasma applicator + intelligent control unit + AI navigation module. Each component designed for surgical integration and real-time feedback.
Our work sits at the intersection of physics-constrained machine learning, surgical robotics, and acoustic simulation. Every product we build traces back to a fundamental scientific question -- and every answer gets published.
Active collaborations with university research partners span continuum mechanics, graph neural network theory, and clinical validation. We believe the best way to build trustworthy surgical AI is to submit it to peer review.
Multiple papers in preparation targeting top venues across robotics, medical imaging, and machine learning.
14 years in medical robotics. Wharton MBA. 29 granted patents. Led the team that achieved FDA clearance and completed first-in-human clinical validation. His work bridges the gap between physics simulation theory and real surgical systems -- from the differential equations to the operating table.
Surgery demands understanding across domains. So does our team. We bring together researchers in computational physics, graph neural networks, acoustic engineering, and clinical surgery -- each contributing a piece of the world model.
We hire for depth and intellectual honesty. The problems we solve don't have textbook answers. They require people who can derive the physics, implement the code, and explain the clinical relevance -- all in the same conversation.
We are building autonomous surgical systems where every robot motion is verified by a world model before execution. No black-box predictions. No statistical shortcuts. Physics in, safety out.
The goal: make expert-level surgery accessible globally. When the machine understands tissue the way a master surgeon does -- through force, sound, and material response -- geography stops being a barrier to care.