| --- |
| license: mit |
| tags: |
| - robotics |
| - diffusion-policy |
| - rdp |
| --- |
| |
| # RDP flipbox single-view deployment checkpoints |
|
|
| Reactive Diffusion Policy trained on `real_single_flexiv_tactar_flipbox`. |
|
|
| ## Deployment files |
|
|
| - `checkpoints/at/latest.ckpt`: final Action Tokenizer checkpoint (epoch 600). |
| - `checkpoints/ldp/latest.ckpt`: final Latent Diffusion Policy checkpoint (epoch 400, global step 21252). |
| - `normalization/normalizer.pkl`: exact LDP normalizer computed during training, including action, pose, gripper, wrench, wrist-image and latent-action fields. |
| - `config/resolved/{at,ldp}/config.yaml`: fully resolved Hydra configurations saved by each training stage. |
| - `config/task/*.yaml`: source task shape and action-index configuration. |
| - `config/deployment_config.yaml`: portable relative checkpoint paths and endpoint input contract. |
|
|
| The observation camera is the single wrist view `wrist_img1`. The policy action has 10 dimensions, selected from the converted 20-dimensional VRR action with indices `[0,1,2,3,4,5,6,7,8,19]`. |
|
|
| Both AT and LDP checkpoints are required by the RDP runtime. |
| When using the resolved LDP Hydra config after download, override its original absolute `at_load_dir` with `checkpoints/at/latest.ckpt` and set the LDP checkpoint path to `checkpoints/ldp/latest.ckpt`. |
|
|
