DVIDIA RESEARCH

A hardware reference we can measure.

A six-joint UR5e reference with a coupled Robotiq 2F-85 gripper. This bench measures authored motion, numerical differences and CPU physics cost. Real hardware accuracy remains unmeasured.

Recorded native physics · schematic link centers
Positions from simulation; geometry and contact surfaces are not rendered.

Timing excludes model loading, settling, rendering, sensing, policy computation, transport and real actuation. The display is a schematic replay; it does not drive a robot.

What changes when we halve the physics step?

These matched trajectory differences test numerical sensitivity in this scene. They do not measure the distance between simulation and reality.

MotionStep comparisonJoint RMSETCP distance RMSEPad-center RMSE

Collision coverage and remaining gates

Serial-specific kinematics, mounting and tool calibration, motor response, sensor age, gripper hysteresis, contact/slip measurements and real task trials are the next evidence gates. The fixed base has no upstream collision proxy. The five-finger assets are audited separately; the two-arm assembly is still to be built.