Ge YanContact labReal-robot reference ↗

Physical intelligence, in contact

Prediction meets the physical world.

Interactive physical model0.00 s · 1×
Ballistic prediction
Preparing the workbench…
Drag to orbit · scroll to zoom
Velocity & predicted interceptionContact points & forces
Ready to run
Nudge at 0.10 s

Prediction gets the hand to the ball. The contact geometry and available friction determine whether it stays there.

Reduced-order mechanics · no learned policy
From the real world

Repair, then return to work

In Flex-π, two arms coordinate to repair a gripper and resume manipulation. Self-repair is one example of the precise, contact-sensitive work a parallel gripper can perform.

My view: general robot intelligence will connect high-level reasoning with fast, contact-aware control. A useful plan must survive contact with the world.

Flex-π · supplied real-robot recording
About the physical model

Flight and contact

The sphere translates and spins in three dimensions. Before contact it follows a ballistic trajectory. The controller predicts its intersection with a horizontal capture plane and moves the hand within task-space acceleration, speed, and travel limits.

Contact comes from the sphere touching finite, tilted finger pads. Spring and damping forces act along each contact normal; Coulomb friction acts tangentially and produces both force and torque. The ball is never attached to the gripper. It can miss, slip, rotate, or fall to the table.

Contact-guided alignment

The bit approaches a finite-clearance guide with a chamfered entrance. Spring and damping reactions at the rim, chamfer, walls, and seat move the tool. Position hold keeps its original biased target. Force-guided execution adjusts that target from lateral contact force; it does not read the true center to correct its position. The contact guide magnifies the tip-center coordinates in millimeters; it represents effective clearance, not a literal section of the screw. The dashed line marks the commanded lateral position; the dot and trail show the actual tip.

Once seated, axial load and jaw friction limit the torque that can reach the screw. Detachment removes torque transmission. Individual threads, handle slip, sensor noise, and full robot dynamics are not resolved.

Scope of the comparison

Fixed execution follows the nominal flight and contact commands. Feedback updates the interception from simulated position and velocity, then reacts to contact. Both use the same finger closure and physical equations.

YAM meshes and six-joint kinematics determine the displayed arm poses. The UMI-style gripper is reconstructed from the recording with an approximately 106 mm flange-to-pinch distance. The model does not simulate full arm dynamics, arm self-collision, or collisions with the palm. Task-space limits and contact parameters are illustrative, not measurements of the hardware.

Automatic disturbances are enabled initially: a lateral ball impulse at 0.10 s or an upward and sideways tool disturbance at 3.00 s. Both execution modes receive the same scheduled disturbance. Disable it to inspect the undisturbed trajectory; manual nudges remain available. Bullet time slows the response to 0.1× and freezes for 1.2 seconds. Repair pauses near peak withdrawal, about 0.11 simulated seconds after the bump; catching pauses immediately. Your selected speed is restored afterward. It changes presentation only. Self-repair opens in contact view.

Neither execution mode runs an LLM or Flex-π. This illustrates physical dependencies; it is not a benchmark of model capabilities.

Robot geometry: YAM / I2RT · Flex-π · Hierarchical robot control