--- license: apache-2.0 task_categories: - robotics tags: - LeRobot - act - fanuc - pick-and-place - real-robot configs: - config_name: default data_files: data/*/*.parquet --- This dataset was created using [LeRobot](https://github.com/huggingface/lerobot). ## Dataset Description Real-robot pick-and-place on a **FANUC CRX-5iA** collaborative arm: a dice is picked up and placed on an empty white block, from four camera views. Recorded by a scripted visual-servo pipeline (not teleoperation), so the demonstrations are consistent and the placements are verified — an episode is only saved if a post-hoc survey confirms the dice actually seated on the block it was aimed at. Prepared for training **task-conditioned ACT**: the action has been narrowed to absolute joint targets, and a per-task one-hot rides in `observation.environment_state`. - **Episodes:** 118 · **Frames:** 83,472 · **FPS:** 30 - **Robot:** `fanuc_crx5ia` · **LeRobot format:** `v3.0` - **Cameras:** 4 × 640×480 (4:3) - **Episode length:** 205 – 865 frames (median 727, ≈24 s) ## The contract Read this before training. Several channels are not what a reader would guess. ### `action` — 7 floats, ABSOLUTE joint targets ``` ['J1', 'J2', 'J3', 'J4', 'J5', 'J6', 'gripper'] ``` `J1`–`J6` are **degrees**; `gripper` is binary. The action at frame *t* is the **next frame's state** — absolute positions, not deltas. Joint space rather than Cartesian pose is deliberate. The `W` Euler channel sits **11.2° from the ±180° wrap** (measured range −168.77…−154.41) where a value jumps 360° and an L1 loss cannot recover, while every joint sits mid-range with 80°+ of headroom. A joint vector is also *one* arm configuration where a Cartesian pose is many, so commanding pose would leave the elbow/wrist branch to the controller. And it matches ACT's origin — ALOHA commands joint positions. ### `observation.state` — 13 floats ``` ['J1', 'J2', 'J3', 'J4', 'J5', 'J6', 'x', 'y', 'z', 'W', 'P', 'R', 'gripper'] ``` Wider than the action on purpose: `x/y/z` (mm) and `W/P/R` (degrees) are the TCP pose, free as *input* and informative, but harmful as prediction targets for the reason above. ### ⚠️ `gripper`: **1.0 means OPEN** Both in `observation.state[12]` and in the action's last channel. This is the recorder's convention and it is the **opposite** of the more common 1.0-means-closed. Feed this bit to a policy that assumes the other convention and it reads "open" as "already holding the part". It is the *commanded* latch, captured before the gripper macro fires, so the command correctly precedes the visible finger motion. ### `observation.environment_state` — task one-hot - `0` — **dice_to_block** — "Pick up the dice and place it on the empty white block." - `1` — **dice_from_board** — "Pick up the dice from the board and place it on the empty white block." This is the per-task embedding input. LeRobot's ACT projects `observation.environment_state` with a `Linear` and appends it to the encoder's 1-D tokens — and a `Linear` applied to a one-hot *is* an embedding lookup, so no fork of LeRobot is needed. Normalize it as **IDENTITY**; standardising a one-hot by column statistics stops it being a one-hot. **The width is frozen into trained weights.** Adding a task changes it and invalidates the checkpoint. ## Known caveats - **Episodes 0–79 have reconstructed states.** The robot's telemetry rides a cache that freezes while a motion command holds the socket, so those episodes originally stepped in one frame per move. They were densified by linear interpolation between observed boundary poses — legitimate here because every move is a straight-line Cartesian move, but validated against ground truth it carries **≈2 mm mean / 8.7 mm p95** position error at grasp depth. Episodes 80–117 are natively dense and are the cleaner subset. Consider validating on 80–117, or weighting them higher. - **Task balance is uneven:** 2 tasks, but the board-recovery task has only 6 episodes against 112 for the main task. - **Video is 4:3 as recorded.** No cropping or letterboxing has been applied. - Images are real throughout; only the state channel was reconstructed. ## Training ```bash python -m lerobot.scripts.lerobot_train \ --dataset.repo_id=azorematter/dice_white_pnp_act_joint7 \ --policy.type=act \ --policy.chunk_size=40 \ --policy.n_action_steps=40 \ --batch_size=8 \ --steps=100000 ``` `chunk_size` is the parameter most worth tuning. 40 steps ≈ 1.33 s of motion at 30 fps — chosen to be about a second (per ACT's own tuning guidance) while still outlasting the ~1.15 s command round trip on this cell, so the arm never stalls waiting for the next chunk. Keep `n_action_steps == chunk_size` so each chunk is executed fully, and leave temporal ensembling **off** — it forces `n_action_steps=1`, which at this cell's command rate is ~14 minutes per episode. Scale `batch_size` to your GPU and the learning rate with it (batch 64 ↔ 5e-5, batch 8 ↔ 1e-5). Four 480×640 views produce ~1200 encoder tokens, so batch 64 will not fit on a single consumer card. ## Dataset Structure [meta/info.json](meta/info.json): ```json { "codebase_version": "v3.0", "fps": 30, "features": { "observation.state": { "dtype": "float32", "shape": [ 13 ], "names": [ "J1", "J2", "J3", "J4", "J5", "J6", "x", "y", "z", "W", "P", "R", "gripper" ] }, "action": { "dtype": "float32", "shape": [ 7 ], "names": [ "J1", "J2", "J3", "J4", "J5", "J6", "gripper" ] }, "observation.images.gripper": { "dtype": "video", "shape": [ 480, 640, 3 ], "names": [ "height", "width", "channels" ], "info": { "video.height": 480, "video.width": 640, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 30, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {}, "is_depth_map": false } }, "observation.images.cam0": { "dtype": "video", "shape": [ 480, 640, 3 ], "names": [ "height", "width", "channels" ], "info": { "video.height": 480, "video.width": 640, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 30, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {}, "is_depth_map": false } }, "observation.images.cam1": { "dtype": "video", "shape": [ 480, 640, 3 ], "names": [ "height", "width", "channels" ], "info": { "video.height": 480, "video.width": 640, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 30, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {}, "is_depth_map": false } }, "observation.images.cam2": { "dtype": "video", "shape": [ 480, 640, 3 ], "names": [ "height", "width", "channels" ], "info": { "video.height": 480, "video.width": 640, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 30, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {}, "is_depth_map": false } }, "timestamp": { "dtype": "float32", "shape": [ 1 ], "names": null }, "frame_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "episode_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "task_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "observation.environment_state": { "dtype": "float32", "shape": [ 2 ], "names": [ "dice_to_block", "dice_from_board" ] } }, "total_episodes": 118, "total_frames": 83472, "total_tasks": 2, "chunks_size": 1000, "data_files_size_in_mb": 100, "video_files_size_in_mb": 200, "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet", "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4", "robot_type": "fanuc_crx5ia", "splits": { "train": "0:118" } } ```