Troiaaa/pi0.5-WSMLo1hknaNw

π0.5 AXIS v1.0 — time-localized first-step noise collapse

Canonical parent: ApexUltron/pi0.5-KX774qZu7mZD @ 46abf397fdb40525841ece710b544b6edcb928a0 (the OpenRoboto competition-6 champion at training time, UID107).

Method. Distillation of the parent's own low-noise sampling behaviour into a policy whose output no longer depends on the sampler's input noise.

1. On-policy rollouts of the parent in the AXIS v1.0 scenes, sampled with a fixed per-task anchor noise (zero for most tasks), with small random action perturbations for state coverage (≈2.3k rollouts, ≈14k policy queries). No evaluation trial addresses or reserved seeds were used.
2. Training objective on the unchanged 10-step Euler sampler: the first Euler step from any noise ε must land on the parent's first-step point from the anchor noise, ‖(ε − 0.1·v_θ(ε, t=1)) − (ε* − 0.1·v_parent(ε*, t=1))‖².
3. Time-localized parameterisation: only the adaRMS time-conditioning kernels (pre_attention_norm_1, pre_ffw_norm_1, final_norm_1 modulation Dense kernels and time_mlp_out) are trained, and every update is projected onto the orthogonal complement of the conditioning vectors at the inference times t = 0.9 … 0.1. The velocity field for t ≤ 0.9 is therefore exactly the parent's (up to BF16 rounding); only the t = 1 step changes.

3,000 AdamW updates, batch 32, LR 1e-4, JAX/openpi, BF16 deployment-precision forward with F32 master weights. All other tensors are byte-identical to the parent; normalization statistics are unchanged. Architecture, tokenizer, action horizon and inference sampler are unchanged.

Local development results (official evaluator, 30 tasks × 20 trials): parent 500/600 on policy seeds 20260907 and 20261101; this model 554/600 and 542/600.