123engineer/pi0.5-owPypCo44VKT

π0.5 AXIS joint-space policy — v1 step 2000

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

Parent and training method

• Parent (canonical round base): Fisher-Wang/pi05-axis-v0.2-all30-40p50 @ 6c1a780d59551ea0a09204a592ead726a809e80b
• Method: full fine-tune (no LoRA) with openpi @ 15a9616a, config pi05_axis_joint (π0.5, discrete state input, action horizon 10), 2000 steps, global batch 256, AdamW, cosine LR 5e-5 → 5e-6, EMA 0.999, training-time joint-state noise (σ 0.005 rad).
• Data: perturbed, checker-filtered MuJoCo 3.11 rollouts of AXIS tasks 31, 33, 34, 35, 40, 41, 46, 48, 51, 52, 53, 54, 56 (2,528 episodes, 142k frames at 5 Hz), rendered from camera0 256×256.
• 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 (rebuilt AXIS sim, 20 trials per task)

Task	This model	Base (official)
31	35%	0%
33	95%	95%
34	40%	0%
35	100%	95%
40	100%	20%
41	95%	85%
46	70%	45%
48	85%	60%
51	100%	100%
52	75%	0%
53	100%	70%
54	95%	80%
56	20%	5%

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.