so101-peg-transfer-arm05-dp

SO-101 arm05 deployed diffusion policy

This is the exact best.ckpt file used by the arm05 inference service. The service-reported SHA-256 was checked against the local file before packaging: 9a170e035305ca2c8937a728abeee5ec014c205a538365367cf57a2291f39d67. It selects epoch 26 and loads the EMA state dictionary. Its deterministic diffusion validation loss is 0.0142956068739295; this metric does not establish physical task success.

Training data: knightnemo/so101-peg-transfer-arm05, 57 episodes / 40,962 frames. Source fingerprint: b655454652031746f2451fdcfb80688dc1a15ae5447ea241ebc798af87950225. The completed training budget and learning curve are retained in config.yaml, metrics.csv, and run_receipt.json; the best checkpoint can be from an earlier epoch than the last training epoch.

The checkpoint uses the custom so101.visual_diffusion_bc.v1 format from Understand_Fleet_RL. It includes training/EMA weights, optimizer, scheduler, scaler, normalization, config, episode split, counters, metrics and RNG state. For inference, reconstruct base_policy.models.dp.vision_policy.VisionDiffusionPolicy from the embedded config/statistics and load checkpoint['ema']. See experiments/so101_dp_deploy.py and docs/so101_diffusion_bc.md for the exact observation preprocessing and serving path; this is not a Transformers checkpoint.

Inputs are front/wrist RGB histories and calibrated LeRobot-normalized joint states. Outputs are absolute joint targets. deployment_metadata.json records the actual observation/prediction/action horizons, control frequency, joint order, and deployed startup pose. Its source model is specific to the recorded arm05 peg-transfer setup.