A learned grip adapted to unfamiliar virtual tissue.
Virtual tissue gives a robot somewhere to practice difficult grips before touching real material.
Teach a gripper to retract soft tissue without crushing it or letting it slip. Vary the tissue properties, then compare the learned controller with a fixed plan on virtual tissue it has never seen.
Edited demonstration briefThe result
The learned grip succeeded more often on unseen virtual tissue
The learned controller handled unfamiliar virtual tissue more successfully than the fixed plan, while still struggling with some stiff, slippery materials.
Simulation only. No real tissue or clinical validation.
Recorded simulation
The same virtual tissue produced different gripping outcomes.
Compare a fixed grip with the learned controller on the same virtual tissue.
Recorded outputs, replayed in your browser. No new simulation is running.
Virtual tissue lets us test grips across materials before physical trials.
The world we created
MuJoCo’s finite-element model computes how tissue deforms as a gripper pulls it. Changing stiffness and friction creates different outcomes, while the simulation records internal strain alongside force, jaw position and camera inputs.
What this enables
A controller learns to adjust its grip from those sensor inputs. A separate fast model predicts strain, deformation and outcomes. The recorded tests include crushing and slipping; the tissue slipped before reaching its modeled tear limit.
Virtual tissues varied in stiffness and friction.
Vary tissue stiffness, friction and grip settings. Record both the sensor readings and the deformation behind them.
The controller learned to adjust its grip from feedback.
Train a controller to respond to sensor feedback, and a separate model to predict hidden strain.
Unseen tissues exposed both successful grips and failures.
Replay the learned controller and fixed plan under the same unseen conditions. Reveal both success and failure.
The same virtual tissue supplies practice, hidden-state targets and a controlled test of whether the robot handles unfamiliar materials better.
Original research artifacts
The original files retain the recorded measurements and model details.