Physical AI Lab 01

When does feedback matter?

Compare a scheduled pick-and-place with a feedback-guarded controller. Same physics. Measured outcomes.

What you will learnDistinguish commanded motion from verified load support; test a one-factor repair.

What runs herePhysX 5.6.1 contacts; designed waypoint controller with optional feedback guards. No learned policy.

Completion criteriaMeasured baseline, diagnostic evidence, controlled repair comparison, then four frozen transfer cases.

Choose a task preset, then run. Checkmarks come from recorded physical evidence, not clicking through.

Interactive workcell

Pick. Place. Verify.

Preparing the workcell…

Starting the physics engine

Loading PhysX and preparing the robot…

01Observe
02Plan
03Act
04Verify

Measured state

Updated from the simulation

Physics engine: Loading
Experiment time: 0.00 s
Center height: —
Object speed: —
Gripper gap: —
Contact state: Waiting

Measured at 120 physics steps/s; graph sampled at 10 Hz. No generated or predicted outcomes.

Open: u(t) = waypoint(schedule, t)
Closed: same u(t); advance only if guard(state)
contact ≥ 8 frames → lift y ≥ 0.85 m → retain
place: |x−xb|,|z−zb| < .27 m; |y−ys| < .055 m
settle: speed < .08 m/s, spin < .2 rad/s for .5 s

One 15 mm closure retry; then abort. μΣN ≳ mg is intuition, not the solver or a success oracle. All bodies remain native PhysX contacts. Occupancy is a user-set safety flag, not person detection.

Download controller source

Run a matched pair for comparable evidence.

Mission control

Set your experiment

Help · step-by-step

Select Measure; run Matched pair. Inspect Events and Sensor traces.
Select Diagnose; choose the measurements that test whether the part is supported, not merely commanded.
Select Correct. Change only friction or mass, keeping the baseline part, bin and scenario. Run closed loop.
When placement is verified, run Matched pair at these settings. Compare both the decisions and actual placement.
Freeze and test. Your one-factor repair is applied to new part, bin and load conditions. The final occupied case must stop safely.
Export JSON history and CSV traces. Reset cancels motion but does not erase the scored attempt.

Controller

Same waypoints, with measured contact/lift/retention/settling guards.

CommandMove the red cube to Bin B.

Generated from your selections.

Object Destination

Scenario

Shared contact friction μ1.2 Part mass (kg)

Same initial poses and settings within each pair. Sweep uses μ = 0.1, 0.6, 1.2.

Which evidence tests load support?

Select Measure and run a matched pair.

Development runs are repeatable. Transfer: one frozen attempt on this device; no LMS grade or tamper-proof certification.

Action checks

Target identified
Destination authorized
Workcell clear
Grasp verified

Outcome

Ready to explore the closed loop.