OpenArm π0.5 BSP action expert, 10 epochs
Final LeRobot checkpoint at step 62,534 (about 10 epochs). Training exited successfully on 2026-09-29. The PaliGemma vision/language backbone was frozen; the BSP action expert and action projections were updated. The policy type is relative_spline_pi05 and requires the matching external lerobot_policy_spline implementation.
The repository root contains the complete pretrained_model export, including its pre/postprocessors and tokenizer. Optimizer state and raw demonstrations are not included. The recorded training dataset is nkmurst/openarm_teleop_udp_d435_merged_20260725_lerobot. This is a research checkpoint; robot success and safety have not been established by the training run.
model.safetensors SHA-256: 11f989cccb53edbaa7fea0a7c1345a1dc3c406675293e9ec0cba251a456cc187.