| # RoboCOIN RIFT inference |
|
|
| Existing final RIFT checkpoint from 10 epochs / 9,340 optimizer steps on object |
| classification (191 episodes), red/yellow test-tube mixing (50), and basket stacking |
| (101). Original source provenance is in [checkpoint_manifest.json](../checkpoint_manifest.json). |
|
|
| ```bash |
| hf download PoopBear/RIFT --include 'robocoin_multitask_mm/rift_10ep/*' 'assets/*' --local-dir ./checkpoints |
| export DIFFSYNTH_MODEL_BASE_PATH=./checkpoints/assets |
| ``` |
|
|
| Load `./checkpoints/robocoin_multitask_mm/rift_10ep/rift_step009340.pt` with the adjacent |
| `config.yaml` and `dataset_stats.json`. The checkpoint includes anticipation/probe |
| parameters; preserve all model state and the packaged processor settings. For full |
| RIFT checkpoint loading, skip separate pretrained video/ActionDiT initialization |
| (`skip_dit_load_from_pretrain=true`, `action_dit_pretrained_path=null`) and enable |
| `load_text_encoder=true` for uncached prompts. Use `weights_only=True` for the policy |
| checkpoint. No raw dataset or training-view archive is needed for live inference. |
|
|
| ## Observation and action contract |
|
|
| - Policy action and proprioceptive state order is |
| `[left_arm (6), left_gripper (1), right_arm (6), right_gripper (1)]` (14 values). |
| The raw dataset's packed order is `[left_arm (6), right_arm (6), left_gripper (1), right_gripper (1)]`; |
| reorder by named fields before policy preprocessing. |
| - Arm joints remain in radians. When starting from the original RoboCOIN source |
| representations, multiply |
| classification and test-tube gripper values by `180/pi` to obtain the policy's |
| millimeter convention; basket gripper values remain unchanged. Apply this to both |
| actions and states, without clipping. Already-converted millimeter values must |
| not be converted again. This conversion follows the original |
| preprocessing's inverse degree-to-radian interpretation; per-subset export logs |
| were not available. For a live robot, map its measured gripper units to the same |
| policy convention, then invert that mapping when sending commands. |
| - Cameras, in order: `head_rgb`, `left_wrist_rgb`, `right_wrist_rgb`. Dataset aliases |
| are `cam_head_left_rgb` for classification/test-tube heads and `cam_high_left_rgb` |
| for basket heads; wrists are `cam_left_wrist_rgb` and `cam_right_wrist_rgb`. |
| - Preserve the packaged 224 x 224 per-camera resizing, three-camera composition, |
| relative-joint transform for both arms, stepwise action normalization, and gripper |
| action normalization `0/100`. Use this checkpoint's statistics for normalization |
| and inverse action processing; do not substitute another dataset's statistics. |
|
|
| The saved config retains loader-compatible fields and original architecture settings. |
| Its normalization path points to the packaged statistics under `./checkpoints/`; |
| relative paths resolve from the working directory. Dataset/cache/output fields are |
| retained metadata rather than inference setup instructions. |
|
|
