Physical AI Lab

Print robosuite composite part slices before concatenating actions; an audited dual-arm example shows wrong ordering swaps targets and named repacking restores geometry.

Use sensor_adr and sensor_dim, rangefinder no-hit semantics and callback-stage order; an actual finite-wall trace separates logging from explicit feedback.

Compare MuJoCo 3.15.0 actuator_force with qfrc_actuator using measured signed tendon gradients, sparse nout-to-nv mapping and post-map limits.

Separate finite surface distance from MuJoCo 3.15.0 margin/gap regimes, explicit-pair parameters and body-exclusion precedence.

Compare a fixed metre delta applied to prescribed position reports or previous desired goals, with a separate absolute target.

A finite graph example separates lane distance, projection bounds and waypoint matching without claiming to run RMF selection.

Five prescribed cubic joins separate path geometry from time-scaled boundary velocity and acceleration, using explicit controls and supplied durations.

A prescribed force at the tool origin separates free couple, lever-arm moment and paired point velocity in a sensor-to-tool wrench transform.

An explicit XYZ convention separates Euler-coordinate dependence from valid tool orientation, using neutral roll/yaw increments and a finite world-X target.

A finite cuboid comparison shows an AABB false positive, a solid overlap and a rejected pair, with a short executable geometry check.