PI0.5 PutCab-CTR-Train50-V4 LoRA 10k
Fresh original PI0.5 JAX base XinY0201/openpi-pi05-base-jax@5e62884fcf8cb8f9fc693c9163ea18d3e3739658, single action expert. Use 10000/ as the OpenPI Orbax checkpoint root; full inference parameter tree includes LoRA and base weights plus normalization assets. Optimizer state retained on training platform.
Dataset Shiki42/PutCab-CTR-Train50-V4@3b9629a729945005dabd7857803218d063134634, all50episodes/14531frames at16.666666666666668FPS. Native14-dimensional absolute state/actions, three RGB views, standard224x224 transforms, padded32-dimensional model actions,horizon50. No idle mask, delta transform or Aloha unit conversion. Prompt: Open the cabinet drawer with the left arm and place the object into it with the right arm.
Two H10080GB,globalbatch16,seed87431,10000updates. PaliGemma LoRA rank/alpha16,expert32; reference filter includes trainable vision/projections. bf16 activations/frozen parameters,float32 trainable parameters. AdamW b1=.9,b2=.95,eps1e-8,weight_decay1e-10,clip1,noEMA. Fixed30k cosine schedule with1000warmup,peak2.5e-5,endpoint2.5e-6,stopped10k.
Real decoder/tokenizer, independent dataset normalization and CPU checkpoint save/reload checks passed before GPU submission. Strict full parameter and optimizer restore verifies finite values and optimizer step10000; decoded-array hashes are retained. No rollout success-rate claim.
Detailed actual runtime/packages/resolved config, source commit+patch, GPU binding, base/dataset manifests, statistics, CPU/save tests and job history are in 10000/experiment/. Platform container NGC PyTorch25.02 runs a separate JAX environment; not claimed byte-identical to historical CTR archived runtime. Job job-b386690e-9dd4-4cad-a49d-c991216be4d1, v2sam-exo2ego-fusion-official24-k15.