| --- |
| license: apache-2.0 |
| task_categories: |
| - robotics |
| tags: |
| - robodojo |
| - annotations |
| --- |
| |
| # RoboDojo progress annotations |
|
|
| Frame-level labels over the official RoboDojo episodes, for the six tasks that |
| require remembering something seen earlier in the episode. They carry no images, |
| states, actions or episodes; the observations and actions a policy trains on |
| come from [`RoboDojo-Benchmark/Robodojo`](https://huggingface.co/datasets/RoboDojo-Benchmark/Robodojo). |
|
|
| Published so that the training recipe behind our RoboDojo entry can be run by |
| anyone: without these labels the keyframe-memory stage has nothing to learn |
| from. |
|
|
| ## What is in it |
|
|
| 600 files, `<task>/trajectory_<n>/progress.json`, covering all 100 episodes of |
| each of the six tasks. Plus `index.json`, which maps each task directory to the |
| official task instruction and states the schema. |
|
|
| | field | meaning | |
| | --- | --- | |
| | `keyframes` | the frames at which the task made visible progress that has to be remembered | |
| | `meta[].frame` / `meta[].box` | `[x1, y1, x2, y2]` in raw-video pixels, the object that progress was about | |
| | `meta[].window` | `[first, last]` — the frames at which recording that keyframe still counts as correct | |
| | `total_frames` | the episode's length, so a file can be checked against the episode it describes | |
|
|
| `meta[].window` exists because a keyframe is a moment, not a frame: several |
| consecutive frames show the same progress, so a selector that answers one frame |
| early is not wrong. Training scores the answer against the whole interval |
| rather than against a single index. |
|
|
| ## Per-task shape |
|
|
| | task | episodes | keyframes each | boxes each | acceptance intervals each | |
| | --- | --- | --- | --- | --- | |
| | `cover_blocks` | 100 | 4 | — | — | |
| | `imitate_sorting_sequence` | 100 | 5 | 5 | — | |
| | `match_and_pick_from_conveyor` | 100 | 1 | — | 1 | |
| | `press_by_number` | 100 | 5–21 | — | — | |
| | `swap_T` | 100 | 1 | — | — | |
| | `swap_blocks` | 100 | 7 | — | 1 | |
|
|
| Where a task has no acceptance interval, the selection answer is scored against |
| the single annotated candidate. |
|
|
| `swap_T` and `match_and_pick_from_conveyor` carry one keyframe, at frame 0: what |
| has to be remembered is the initial position and orientation of the two blocks, |
| and the first object on the conveyor. Both are visible at the start. |
|
|
| ## Checking them against the official episodes |
|
|
| An annotation bound to the wrong episode would train a policy to recall the |
| wrong frames. Each file therefore records its episode's length; the loader binds |
| by task instruction and stops if a length disagrees. |
|
|
| ## How they were produced |
|
|
| For each task we decided by hand what counts as a keyframe — the moments at |
| which the task makes the progress a policy has to remember, such as a block |
| being covered or a button press being committed. A script then located the |
| frames that satisfy the rule in each episode, and every file was checked by a |
| person and corrected where the script had put a frame in the wrong place. |
|
|
| ## Using them |
|
|
| ```bash |
| hf download x-square-robot/robodojo-memory-annotations --repo-type dataset \ |
| --local-dir <bench_root>/data/robodojo_memory_annotations |
| ``` |
|
|
| The training code that reads them is in |
| [wall-x](https://github.com/X-Square-Robot/wall-x) |
| (`wall_x/data/backends/lerobot/robodojo/`); see the WallX policy README in XPolicyLab |
| for the two training stages. |
|
|
