π0.5 AXIS joint-space policy — v6b step 750

OpenRoboto (Bittensor SN80) AXIS v1.0 · π0.5 submission.

Parent and training method

  • Parent (current season champion; king-of-the-hill rules allow building on it): pfenzi/pi0.5-8wKVKfMLCNbx @ a4a5f5711261b661c2fde6fa2f067acc9a09b050
  • Method: full fine-tune (no LoRA) with openpi @ 15a9616a, config pi05_axis_joint (Ï€0.5, discrete state input, action horizon 10); this export is step 750. Global batch 24, AdamW (clip 1.0), flat LR 1.5e-6, EMA 0.999, 1,250 steps configured; no state-noise augmentation.
  • Data: rollouts generated in the official AXIS v1.0 runtime (openroboto-evaluation @ a1124074) on all 30 tasks: the parent's own checker-passing rollouts (clean and with small executed-target noise) labelled with its proposals, plus checker-passing rollouts of two public submissions on tasks where they beat the parent (MechaTrainer/pi0.5-rneQYn9nFopV @ f0a10607 on 34 and 43; jenny08311/pi0.5-TgbWfarpxG8Q @ 9ea0049e on 42), and for tasks where it verifiably helped, parent rollouts with a small time-indexed correction added to its proposals (labels = executed targets, 6.5% of frames per task). About 150k frames at 5 Hz.
  • Interface unchanged: camera0 RGB + 9D joint state in, 9D absolute joint targets out; assets/axis-v0.1-task501-runtime-v1/norm_stats.json is byte-identical to the parent's.

Local evaluation (official AXIS v1.0 runtime, openroboto-evaluation @ a1124074)

Mean over 30 tasks: this model 0.822, parent 0.813 (trials in brackets).

Task This model Parent
22 100% (20) 95% (20)
31 60% (20) 47% (60)
33 100% (20) 100% (20)
34 35% (20) 27% (60)
35 100% (20) 100% (20)
37 50% (20) 43% (60)
40 100% (20) 100% (20)
41 90% (20) 95% (20)
42 80% (20) 60% (60)
43 60% (20) 70% (60)
44 55% (20) 87% (60)
46 100% (20) 100% (20)
48 95% (20) 90% (20)
49 95% (20) 95% (60)
50 45% (20) 58% (60)
51 100% (20) 95% (20)
52 95% (20) 90% (20)
53 100% (20) 100% (20)
54 100% (20) 95% (20)
55 85% (20) 67% (60)
56 95% (20) 95% (20)
57 100% (20) 100% (20)
501 40% (20) 43% (60)
502 65% (20) 62% (60)
503 95% (20) 95% (60)
504 85% (20) 88% (60)
505 90% (20) 93% (60)
506 95% (20) 95% (20)
514 100% (20) 83% (60)
757 55% (20) 70% (60)

License

Derived from π0.5 (openpi) and Gemma: see LICENSE_GEMMA.txt (Gemma Terms of Use, including the Section 3.2 use restrictions), LICENSE_OPENPI.txt and NOTICE.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Model tree for 123engineer/pi0.5-v6bs750-owPypCo44VKT

Finetuned
(2)
this model