bluecube_101_0715_policy

ACT policy trained on JunJunJunya/bluecube_101_0715 alone (100 episodes), lerobot ACT defaults (optimizer_lr=1e-5, optimizer_lr_backbone=1e-5, batch_size=16, no image augmentation), for 20,000 steps.

Training completed all 20,000 steps, but the local disk ran out of space while writing the final checkpoint's optimizer_state.safetensors (resume-only state), so this upload is the model weights only (pretrained_model/) -- fine for inference, just not resumable as a training run.

Trained locally via Training_local.py in this repo.