PI0.5 scan object concurrent LoRA 10k
Fresh original PI0.5 JAX base XinY0201/openpi-pi05-base-jax@5e62884fcf8cb8f9fc693c9163ea18d3e3739658, standard single action expert. Use 10000/ as the OpenPI Orbax checkpoint root. It contains the full base+LoRA inference parameter tree and normalization assets; optimizer state retained on the training platform.
Dataset Shiki42/ctr-scan-object-concurrent-20260911@2419a110014733297689002a04e77faf1c438300, all50episodes/11251frames at25FPS. Native14-dimensional absolute actions/state; three RGB cameras; standard224x224 preprocessing; pad32 and horizon50. No delta/Aloha unit conversion or idle mask. Native prompt: Scan the object.
Two H10080GB;globalbatch16;seed87431;10000updates. PaliGemma LoRA rank/alpha16,expert32, reference filter includes trainable vision/projections. bf16 activations/frozen weights,float32 trainable weights. AdamW b1=.9,b2=.95,eps1e-8,weight_decay1e-10,clip1,noEMA. Fixed30k cosine LR plan with1000warmup,peak2.5e-5,end2.5e-6,stopped10k.
Dataset-specific full statistics, real decoder/tokenizer and CPU save/reload gates passed before GPU submission. Final parameter and optimizer trees strictly restored and checked finite, optimizer step10000 verified; hashes and receipts included. No rollout success-rate claim.
10000/experiment/ records actual runtime/package inventory, resolved config, source commit/patch, GPU binding, pinned dataset/base manifests, normalization/gates/compatibility and history. Platform NGC PyTorch25.02 executes a separately built JAX runtime, not claimed identical to old archived CTR runtime. Job job-f279f335-17e5-49ea-a29c-ce48637aacec, v2sam-exo2ego-fusion-official24-k23.