Updated 24 September 2026. A privileged scripted UR5 + RG2 expert empties a free 16 g Hazel box of 15 hollow PLA tubes (20 mm × 80 mm, 10 g, piled three layers deep), dropping each tube into either side box. Tubes are only grasped across their axis: the fingers dig into gaps between tubes, close while descending, follow the tube as it moves, and retry deeper after a miss. Colour sorting is on hold: picking from a pile was where it failed, so the picking task is being solved first.
| Tubes at start | Episodes | Picked | Picked rate | Fully cleared | Valid | Mean time | Time cap | Timed out |
|---|---|---|---|---|---|---|---|---|
| 14 | 2 | 25/28 | 89% | 1/2 | 0/2 | 312 s | 735 s | 0/2 |
| 15 | 126 | 1704/1890 | 90% | 63/126 | 0/126 | 396 s | 735 s | 1/126 |
A tube counts as picked once it is out of the source box and released. Valid means the robot never hit the table or room (touching and pushing the boxes is allowed), no box toppled (tilt past ~25°), robot penetration ≤ 3 mm, box contacts ≤ 0.8 mm, tube contacts ≤ 1.5 mm, joint speed ≤ 1.57 rad/s and finite physics. Tubes left in the box: 109 walled in (no side grasp clears the box wall and no roll freed them), 74 missed three grasps in a row, 0 never given up on (time ran out first). Per-episode results · Summary
Before the dig-in grasps, live re-aim and CG-80 contacts; that version picked 49% of tubes over 1 to 10 tube piles.
| Physics | MuJoCo Warp, 1 ms step, 2 ms contact time constant, CG solver 80 iterations, elliptic cones |
| Collision geometry | box = floor plate + 4 tapered wall plates (visual mesh unchanged); tube = one capsule (hollow visual mesh) |
| Contacts per settled world | ~35 (was ~1,400 with 177-piece box shells and 32-sector tubes) |
| Throughput, physics only | 366 env-steps/s at 1,024 worlds (was 9 env-steps/s at 32 worlds) |
| Memory sizing | njmax 1024 rows/world, nconmax 128/world, naccdmax 32/world (was up to 17 GB at 512 worlds) |
| Non-finite (NaN) worlds, expert acting | 0 / 256 over 90 s; 0 in the 128-episode 15-tube evaluation |
| Box topples, 64 active 15-tube worlds, 125 s | 0 with CG-80 (16 with CG-40: unconverged impulses launched free boxes; Newton: 0 topples but 6 NaN) |
| Controller | 10 mm and 3 deg per step: 146 mm/s TCP, 44 deg/s wrist, 0.6 rad/s peak joint speed |
Fifteen tubes cover more than the 149 × 94 mm box floor, so the bottom layer is packed wall to wall, and a side grasp needs a corridor about 60 mm wide across the tube. The expert takes the top layers first, digs a finger into the gap next to a crowded tube so the descent pushes it free, and rolls or slides tubes that are still walled in. What remains are tubes pinned against the long walls under their neighbours and grasps that miss three times. Every episode still exceeds the 0.8 mm box-penetration gate: pressing into a three-layer pile pushes the box floor about 1 mm into the table. The earlier colour-sorting work and its videos remain on the tube sorting page.