PI0.5 PutCab-Sequential-Train50-V4 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. The complete inference parameter tree includes base and LoRA weights, plus normalization assets. Optimizer state remains on the training platform.
Dataset Shiki42/PutCab-Sequential-Train50-V4 at c6022cf9164b34c7244dfede8ece4d20a999ab82; all 50 episodes/18143 frames at 16.666666666666668 FPS, native14-dimensional absolute robot state/actions, no delta/Aloha unit conversion or idle mask. Three native RGB views are mapped by the attached training_config.py and resized to224x224; state/actions padded to32; horizon50. Prompt: Open the cabinet drawer with the left arm and place the object into it with the right arm.
Two H10080GB, globalbatch16,seed87431,10,000updates. PaliGemma LoRA rank/alpha16, action expert32, reference freeze filter including trainable vision/projections. bf16 activations/frozen weights and float32 trainable weights. AdamW b1=.9,b2=.95,eps1e-8,weight_decay1e-10,clip1;noEMA. Fixed30k cosine LR curve with1000warmup,peak2.5e-5,endpoint2.5e-6; training stops10k.
Independent dataset statistics and real reader/tokenizer checks passed before GPU allocation. Normal checkpoint saving uses the previously qualified Orbax compatibility fix. Every parameter and optimizer array is strictly restored, checked finite, and optimizer step10000 verified; decoded-array hashes are recorded. No closed-loop rollout success-rate claim is made.
Detailed runtime/package versions, resolved config, source commit/patch against pinned upstream, GPU binding, dataset/base file manifests, normalization, CPU/save gates, compatibility patches and job history are in 10000/experiment/. Platform container NGC PyTorch25.02 runs a separately built JAX environment. This is not claimed byte-identical to historical CTR archived runtime. Training identity: v2sam-exo2ego-fusion-official24-k9-r2, job-64e89605-fc36-485b-8708-8b7a7e488c1e.