---
pretty_name: DexJoCo Delta Actions for DICE-RL
tags:
  - robotics
  - reinforcement-learning
  - dexjoco
  - dice-rl
---
# DexJoCo delta-action data for DICE-RL

Processed image-and-state demonstrations. Action layout: world position delta (3),
world relative SO(3) rotation vector (3), absolute Allegro joint targets (16).
This is 22 values for single-arm tasks; bimanual tasks concatenate right then
left arm blocks for 44 values, preserving each recorded hand's joint ordering.
Deltas use the measured observation at the same timestep. Both states and actions
are min-max normalized; BC and RL share the same normalization statistics.
Observation version: `initial_object_pose_v1`.

| Task | Episodes | Transitions | State dimension | Action dimension |
| --- | ---: | ---: | ---: | ---: |
| pinch_tongs | 100 | 40469 | 31 | 22 |
| pick_bucket | 100 | 43581 | 38 | 22 |
| hammer_nail | 100 | 21563 | 38 | 22 |
| click_mouse | 100 | 32702 | 31 | 22 |
| fold_glasses | 100 | 53878 | 38 | 22 |
| water_plant | 100 | 27745 | 38 | 22 |
| bimanual_assembly | 100 | 54630 | 61 | 44 |
| bimanual_hanoi | 100 | 107764 | 50 | 44 |
| bimanual_microwave_cook | 100 | 68503 | 61 | 44 |
| bimanual_photograph | 100 | 34377 | 61 | 44 |
| bimanual_unlock_ipad | 100 | 41042 | 61 | 44 |

Each task contains `ph_pretrain/train.npz`, `ph_finetune/train.npz`, their
`normalization.npz` files, and `episodes.json`. The manifest records SHA-256 hashes.
RL rewards and terminals are 1 only at each retained demonstration's final frame.
These are synthetic end-of-demo success labels, not replay-verified success times
for the original recordings. No model checkpoints are included.

## RGB images (all native cameras)

Single-arm episodes include `front`, `wrist`, `ego_left`, and `ego_right`.
Bimanual episodes include `ego`, `wrist_left`, and `wrist_right`.
The original MP4 files are preserved byte for byte, including their resolution.
`images.json` maps each processed timestep to its exact source video frame;
the same initial-prefix cut is used for states, delta actions, and images.
Each task stores videos under `videos/episode_XXXX/<camera>.mp4`.
On the Hub, `videos.zip` bundles every camera video for a task. The download
script automatically extracts and verifies them, preserving the original bytes.

Use `agent.dataset.dexjoco_image.DexJoCoImageDataset` (BC) or
`DexJoCoImageQLearningDataset` (offline RL) from the accompanying checkout.
The loader returns RGB uint8 `(B, history, 3 * cameras, height, width)` in
`conditions["rgb"]` and normalized robot proprioception (first 23 state columns
for single-arm tasks, first 46 for bimanual tasks)
in `conditions["state"]`. Task/object privileged state is excluded by default.
Frames are decoded on demand, with bounded caches and exact PTS-based seeking.
Choose cameras and image size at training time; defaults are `front,wrist` for
single-arm tasks, all three cameras for bimanual tasks, and 224 x 224 pixels.
The loader reads the published videos directly; no source-dataset
download or offline image conversion is required.

Choose the number of published episodes used for training with
`max_n_episodes`; all native camera views remain available.


## Array format

Load NPZ files with `numpy.load(path, allow_pickle=False)`.

| Key | Shape | Meaning |
| --- | --- | --- |
| `states` | `(T, obs_dim)` | Normalized observations, float32 |
| `actions` | `(T, action_dim)` | Normalized pose deltas and absolute hand targets, float32 |
| `traj_lengths` | `(N,)` | Episode lengths, int64; sum equals T |
| `rewards`, `terminals` | `(T,)` | Finetuning split only, float32 |

`normalization.npz` stores `obs_min`, `obs_max`, `action_min`, `action_max`,
`action_mode`, `normalized`, and `observation_version`.
Normalization is `2 * (x - minimum) / (maximum - minimum + 1e-6) - 1`.
Unnormalize actions before decoding them with the latest measured robot pose:
`target_xyz = observed_xyz + delta_xyz` and
`R_target = Exp(delta_rotvec) @ R_observed`. Hand targets stay absolute.
The state and raw target quaternions use wxyz ordering.

Bimanual state begins `[right_pose7, left_pose7, right_hand16, left_hand16]`.
Bimanual raw targets and `decode_actions` output instead use
`[right_pose7, right_hand16, left_pose7, left_hand16]`.
Before sending decoded targets to the simulator's policy-mode API, reorder
them to `[right_pose7, left_pose7, right_hand16, left_hand16]`.

For `initial_object_pose_v1`, object pose features describe the initial object
poses, matching the original recordings. They do not track moving objects.

Use the accompanying DICE-RL checkout's `bash script/download_dexjoco_hf.sh`.
See its `README_DEXJOCO.md` for installation and action decoding.
Source recordings: [DexJoCo](https://huggingface.co/datasets/DexJoCo/DexJoCo-Datasets-Raw).