assembly-lab
Learned insertion of an ECU-sized part into a locating fixture on a simulated Franka Emika Panda (MuJoCo), with a wrist force-torque sensor. 1 mm clearance per side, 2 mm lead-in chamfer. Code: github.com/pavanyadava007/assembly-lab
Simulation only, no real robot. Policies are trained on 479 scripted demonstrations (104,848 frames), not human teleoperation. Every number below is read from the evaluation files: 200 unseen seeds per cell, Wilson 95% intervals.
Insertion success
| Policy | Nominal perception | Fixture perceived 2 mm off |
|---|---|---|
| Scripted expert (reference) | 200/200 (98-100%) | 165/200 (77-87%) |
| Scripted expert, Cartesian impedance control | 200/200 (98-100%) | 194/200 (94-99%) |
| MLP behavior cloning | 179/200 (84-93%) | 84/200 (35-49%) |
| MLP behavior cloning, no wrist F/T | 156/200 (72-83%) | 47/200 (18-30%) |
| MLP behavior cloning under impedance control | 51/200 (20-32%) | 38/200 (14-25%) |
| ACT (LeRobot) | 9/200 (2-8%) | 8/200 (2-8%) |
| Diffusion Policy (LeRobot) | 75/200 (31-44%) | 63/200 (25-38%) |
| Diffusion Policy, no wrist F/T | 55/200 (22-34%) | 52/200 (20-32%) |
Does wrist force-torque sensing matter?
Paired on the same seeds, exact McNemar test.
| Comparison | Condition | Only with F/T succeeded | Only without F/T succeeded | p |
|---|---|---|---|---|
| MLP behavior cloning vs without F/T | nominal | 36 | 13 | 1.4e-03 |
| MLP behavior cloning vs without F/T | 2 mm perception error | 57 | 20 | 2.9e-05 |
| Diffusion Policy (LeRobot) vs without F/T | nominal | 39 | 19 | 1.2e-02 |
| Diffusion Policy (LeRobot) vs without F/T | 2 mm perception error | 40 | 29 | 2.3e-01 |