| library_name: lerobot | |
| tags: | |
| - robotics | |
| - imitation-learning | |
| - diffusion-policy | |
| - lerobot | |
| - franka | |
| - visuomotor-policy | |
| # FR3 Cube Full10K Diffusion Policy | |
| Evaluation-ready LeRobot Diffusion Policy checkpoints from cumulative optimizer step 10K through 50K at 2.5K intervals, plus the 60K checkpoint. | |
| ## Policy | |
| - Observation horizon: 2 | |
| - Prediction horizon: 16 | |
| - Action chunk (`n_action_steps`): 8 | |
| - Three native 320×180 RGB views + 22D robot state | |
| - 7D relative Cartesian OSC action | |
| - Conditional 1D U-Net with ResNet18 visual encoder | |
| Each checkpoint is a LeRobot `pretrained_model/` directory containing the model config, preprocessing/postprocessing state, and `model.safetensors`. Load the directory directly with `DiffusionPolicy.from_pretrained(...)` or the UWLab rollout evaluator. | |
| ## Checkpoint provenance | |
| - 10K–27.5K: original run [`7193up9q`](https://wandb.ai/AIDAS-Diffusion/isaaclab/runs/7193up9q) | |
| - 30K–60K: continuation from the 27.5K weights, run [`7ftqgx5v`](https://wandb.ai/AIDAS-Diffusion/fr3_cube_rgb_distillation/runs/7ftqgx5v) | |
| The continuation restored model weights exactly, but the original W&B artifact did not contain optimizer/scheduler state. Therefore these releases are evaluation-ready model checkpoints, not full-state training-resume checkpoints. The cumulative step is encoded in each directory name. | |
| ## Download example | |
| ```bash | |
| hf download jaeikkim/fr3-cube-full10k-diffusion-policy checkpoints/step_050000/pretrained_model --local-dir ./fr3_cube_dpu | |
| ``` | |
| Dataset: [FR3 Cube Full-Episode RL-Teacher 10K RGB](https://huggingface.co/datasets/jaeikkim/fr3-cube-full-episode-rlteacher-10k-rgb). Verify model files with `SHA256SUMS` or `checkpoint_manifest.json`. | |