| --- |
| license: mit |
| library_name: robomimic |
| tags: |
| - robotics |
| - doosan |
| - behavior-cloning |
| - diffusion-policy |
| - vision |
| - imitation-learning |
| --- |
| |
| # Checkpoints |
|
|
| This folder contains selected robomimic checkpoints that were manually copied back |
| to the Doosan workstation for inference tests. |
|
|
| Current inference notes: |
| - recommended path on the robot: `action_mode:=vel_cmd` |
| - pass the same `offset_file` used during data collection |
| - for full-task `open_fridge`, `freeze_target_on_episode_start:=true` is recommended |
| - for vision-only checkpoints, use `bc_robomimic_vis` / `bc_robomimic_vis_runner` and no `offset_file` |
| - for diffusion vision checkpoints, use `diffusion_policy_vis` / `diffusion_policy_vis_runner` with the sidecar `action_stats.json` |
|
|
| | File | Model | Action target | Dataset | Num Episodes | Num Epoche | Notes | |
| |---|---|---|---|---:|---:|---| |
| | `model_epoch_183_best_validation_0.13749891743063927.pth` | BC-RNN State normal | `action.vel_cmd` | `dataset_v1` | 25 | 200 | First 25 episodes of `dataset_v1`. | |
| | `model_epoch_177_best_validation_0.1074166551232338.pth` | BC-RNN State normal | `action.vel_cmd` | `dataset_v1` | 50 | 200 | First 50 episodes of `dataset_v1`. | |
| | `model_epoch_139_best_validation_0.13213149458169937.pth` | BC-RNN State normal | `action.vel_cmd` | `dataset_v1` | 100 | 200 | 100 episodes of `dataset_v1`. | |
| | `model_epoch_177_best_validation_0.11254111714661122.pth` | BC-RNN State normal | `action.vel_cmd` | `dataset_v1_paolo_50` | 50 | 200 | Last 50 episodes of `dataset_v1`. | |
| | `model_epoch_390_best_validation_0.09170645512640477.pth` | BC-RNN State big | `action.vel_cmd` | `dataset_v1` | 100 | 500 | 100 episodes of `dataset_v1`, big model, 500 epochs. | |
| | `model_epoch_390_best_validation_0.09260034952312708.pth` | BC-RNN State big | `action.vel_cmd` | `dataset_v1_paolo_50` | 50 | 500 | Paolo 50 episodes with synthetic augmentation up to 100 episodes. | |
| | `model_epoch_186_best_validation_0.1702012188732624.pth` | BC-RNN Vision normal | `action.vel_cmd` | `dataset_v1` | 50 | 200 | Vision-only BC-RNN on first 50 episodes of `dataset_v1`; test with `bc_robomimic_vis_runner`. | |
| | `model_epoch_327_best_validation_0.16748503260314465.pth` | BC-RNN Vision big | `action.vel_cmd` | `dataset_v2` | 100 | 500 | Vision-only BC-RNN on first 50 episodes of `dataset_v1` + 50 episodes of `dataset_v2`. | |
| | `model_epoch_449_best_validation_0.14368683397769927.pth` | BC-RNN Vision normal | `action.vel_cmd` | `dataset_v2` | 150 | 500 | Vision-only BC-RNN on 150 episodes of `dataset_v2`, standard model. | |
| | `model_epoch_357_best_validation_0.12847926169633866.pth` | BC-RNN Vision big | `action.vel_cmd` | `dataset_v2` | 150 | 600 | Vision-only BC-RNN on 150 episodes of `dataset_v2`, big model. | |
| | `model_epoch_178_best_validation_0.012047586403787136.pth` | Diffusion Policy Vision normal | `action.vel_cmd` | `dataset_v2` | 150 | 600 | Vision-only Diffusion Policy on 150 episodes of `dataset_v2`; inference validated with `diffusion_policy_vis_runner` and sidecar `action_stats.json`. | |
|
|
