Physical AI Lab

The generic interface kept its pilot clearance and mounting-hole pattern at 0 and 180 degrees, but the half-turn exchanged the measured aperture pairings.

Two generic 18 mm slides produced 15.588 mm and 9.000 mm of normal approach, showing why compliance direction matters before force tuning.

The same grasp and final pose produced different intermediate clearances: turning low obstructed the bin, while lifting first cleared every sampled pose.

Two Blender slider rows show why a half-speed reference can be 60 mm behind an unscaled schedule without that difference being tracking error.

Four measured Blender slider cases separate sampled path error from endpoint goal error and show why a later endpoint cannot rescue an earlier path-tolerance abort.

An Isaac Lab demonstration can load and step while state agreement or task success remains unresolved. Use this evidence ladder to diagnose replay results.

MoveIt parallel planning needs a declared selection criterion. See why early stopping, shortest joint-space paths and finite obstacle clearance answer different questions.

A costmap cell with value 255 follows its own unknown-handling branch. Trace the emitted centre point separately from visibility and finite obstacle geometry.

A fixed pad correction aligned one authored linkage pose but left a 12.124-degree pair error at another. Compare finite face angles, edge gaps and center apertures.

Trace robosuite’s componentwise action map before changing gains: asymmetric output bounds can shift zero input, while clipping can collapse distinct commands.