Updated 2026-09-18. Tubes start as a random heap of lying cylinders in the centre basket; the scripted expert sorts red to the left basket and blue to the right using physical grasps only (no teleporting or pinning of tubes). Every number below is read from the run's results.json.
4 of 4 worlds fully sorted, 4 of 4 physically valid (no wall contact above 0.05 N, pad penetration under 3 mm).
| world | tubes | sorted | time | valid | max pad penetration | wall contact |
|---|---|---|---|---|---|---|
| 0 | 4 | 4/4 | 48 s | yes | 0.4 mm | no |
| 1 | 4 | 4/4 | 54 s | yes | 0.5 mm | no |
| 2 | 4 | 4/4 | 50 s | yes | 0.4 mm | no |
| 3 | 4 | 4/4 | 53 s | yes | 0.4 mm | no |
40 of 40 tubes sorted across 4 worlds; 4 of 4 worlds fully sorted; 1 of 4 pass the strict validity gate. Wall-hugging tubes are first nudged off the wall with the closed pads, then grasped.
| world | tubes | sorted | time | valid | max pad penetration | wall contact |
|---|---|---|---|---|---|---|
| 0 | 10 | 10/10 | 146 s | no | 3.5 mm | yes |
| 1 | 10 | 10/10 | 150 s | no | 3.0 mm | no |
| 2 | 10 | 10/10 | 166 s | no | 3.2 mm | yes |
| 3 | 10 | 10/10 | 153 s | yes | 2.7 mm | no |
A world counts as settled when, with the arm still for 5 s, in the last second no tube leaves the basket, none moves faster than 20 mm/s or drifts more than 3 mm, and no tube contact penetrates more than 3 mm. Baseline (previous physics): 22/72 worlds settled with tube penetration up to 5 cm. Current physics and layout: 66/72 worlds settled, 0 tubes ejected during the drop.
| case | settled | ejected | tail speed max | tail penetration max |
|---|---|---|---|---|
| nominal-4 | 8/8 | 0 | 0.2 mm/s | 0.1 mm |
| nominal-10 | 8/8 | 0 | 1.9 mm/s | 0.2 mm |
| dr-4 | 6/8 | 0 | 33.3 mm/s | 0.8 mm |
| dr-7 | 8/8 | 0 | 7.1 mm/s | 0.2 mm |
| dr-10 | 7/8 | 0 | 244.7 mm/s | 0.4 mm |
| fat-long-heavy-10 | 6/8 | 0 | 271.4 mm/s | 0.7 mm |
| thin-short-light-10 | 8/8 | 0 | 2.8 mm/s | 0.2 mm |
| small-basket-7 | 8/8 | 0 | 13.3 mm/s | 0.2 mm |
| wide-basket-10 | 7/8 | 0 | 71.9 mm/s | 0.2 mm |
Settling clip (ten tubes, full randomization):
What changed: capsule colliders instead of cylinders (cylinder pairs get one contact point in MuJoCo Warp and rock), contact time constant 0.04 s to 0.01 s (steady-state penetration under gravity fell from ~1.6 cm to ~1 mm), zero collision margin, zero free-joint damping and armature (the old values froze the tubes' rotation), rolling friction, non-overlapping layered spawn, and a 380×128 mm centre basket so the RG2 fingers have room. Full report: stability README, matrix JSON.
The ACT policy videos and the earlier "fallen pile" clips previously on this page were produced with the unstable physics and, in one case, a scripted grasp lock. They are kept for provenance only: ACT world 0, tipped-pile 4/4 (small tilt), 90° pile (0/4, unstable). The ACT policy has not been retrained on the new task yet.
Export README · Control contract · Manifest · Old ACT evaluation summary