--- license: apache-2.0 tags: - robotics - lerobot - manipulation - imitation-learning - vision-language-action - embodied-ai - teleoperation - so_follower task_categories: - robotics language: - en size_categories: - 1M ## Overview NIAT10 gathers every manipulation dataset recorded by the NIAT Physical AI group into one place, each one preserved in its original [LeRobot](https://github.com/huggingface/lerobot) layout inside its own top-level folder. Rather than a single merged dataset, it is a **curated collection** — you can train on one folder, a task family, or all of them. The collection covers five task families on a shared tabletop setup: pick and place, sorting, stacking, object reorientation, and non-prehensile pushing. Every episode was teleoperated on an **SO-101 (leader–follower pair)** through a leader arm and recorded from an overhead and a wrist camera at 640 × 480 / 30 fps. ## Dataset Statistics | Metric | Value | |---|---| | **Total datasets** | 25 | | **Total episodes** | 2,138 | | **Total frames** | 1,021,300 | | **Recorded duration** | 9.46 hours (09:27:23) | | **Video footage** | 18.91 hours across all camera streams | | **Camera frames** | 2,042,600 | | **Task families** | 5 | | **Robot embodiments** | 1 | | **Average duration/dataset** | 0.38 hours | | **Datasets passing review** | 23 of 25 | **Recorded duration** is wall-clock robot interaction time, computed as `total_frames / fps` from each dataset's `meta/info.json`, so it reflects the current contents of each repo after episode pruning. **Video footage** counts each synchronised camera stream separately — it is the volume of video you will download and decode, not additional robot time. `pick_place_cubes_para` re-annotates episodes that already appear elsewhere in the collection rather than adding new recordings (see [Language variants](#language-variants)). Excluding it, the collection holds **8.60 hours** of distinct recordings across **1,838 episodes**. ## Task Distribution | Task family | Datasets | Episodes | Duration | % of data | |---|---|---|---|---| | Pick and place | 14 | 1210 | 05:01:17 | 53.1% | | Sorting | 4 | 374 | 02:08:26 | 22.6% | | Stacking | 3 | 212 | 00:56:42 | 10.0% | | Non-prehensile push | 2 | 244 | 00:49:07 | 8.7% | | Orientation | 2 | 98 | 00:31:52 | 5.6% | ## Robot Types | Robot type | Datasets | % | |---|---|---| | so_follower | 25 | 100.0% | Recorded with LeRobot v3.0 (25 datasets). Frame rates present: 30 fps (25). ## Contributors Recorded and published under the accounts below. | Account | Datasets | % | |---|---|---| | [NIATphysicalAI](https://huggingface.co/NIATphysicalAI) | 20 | 80.0% | | [Bradx86](https://huggingface.co/Bradx86) | 5 | 20.0% | Please credit the individual contributors above, not only the publishing accounts, when using this collection. ## Contents | # | Folder | Task | Episodes | Frames | Duration | FPS | Review | Source | |---|---|---|---|---|---|---|---|---| | 1 | `Eraser_drawer` | Pick and place | 100 | 79,890 | 00:44:23 | 30 | fix required | [Eraser_drawer](https://huggingface.co/datasets/NIATphysicalAI/Eraser_drawer) | | 2 | `Battery_sort` | Sorting | 106 | 93,878 | 00:52:09 | 30 | OK | [Battery_sort](https://huggingface.co/datasets/NIATphysicalAI/Battery_sort) | | 3 | `Battery_sort_v1` | Sorting | 7 | 5,681 | 00:03:09 | 30 | OK | [Battery_sort_v1](https://huggingface.co/datasets/NIATphysicalAI/Battery_sort_v1) | | 4 | `drawer_cube_screwdriver` | Pick and place | 101 | 54,895 | 00:30:30 | 30 | OK | [drawer_cube_screwdriver](https://huggingface.co/datasets/NIATphysicalAI/drawer_cube_screwdriver) | | 5 | `Drawer_screwdriver_v2` | Pick and place | 51 | 30,709 | 00:17:04 | 30 | OK | [Drawer_screwdriver_v2](https://huggingface.co/datasets/NIATphysicalAI/Drawer_screwdriver_v2) | | 6 | `Drawer_screwdriver` | Pick and place | 34 | 23,221 | 00:12:54 | 30 | OK | [Drawer_screwdriver](https://huggingface.co/datasets/NIATphysicalAI/Drawer_screwdriver) | | 7 | `screwdriver_box` | Pick and place | 47 | 27,536 | 00:15:18 | 30 | OK | [screwdriver_box](https://huggingface.co/datasets/NIATphysicalAI/screwdriver_box) | | 8 | `screwdriver_box_v1` | Pick and place | 20 | 13,829 | 00:07:41 | 30 | OK | [screwdriver_box_v1](https://huggingface.co/datasets/NIATphysicalAI/screwdriver_box_v1) | | 9 | `screwdriver_box_v0` | Pick and place | 10 | 7,498 | 00:04:10 | 30 | OK | [screwdriver_box_v0](https://huggingface.co/datasets/NIATphysicalAI/screwdriver_box_v0) | | 10 | `stack_cubes` | Stacking | 58 | 30,806 | 00:17:07 | 30 | OK | [stack_cubes](https://huggingface.co/datasets/NIATphysicalAI/stack_cubes) | | 11 | `Cup_stack` | Stacking | 100 | 45,646 | 00:25:22 | 30 | OK | [Cup_stack](https://huggingface.co/datasets/NIATphysicalAI/Cup_stack) | | 12 | `eraser` | Pick and place | 60 | 26,893 | 00:14:56 | 30 | OK | [eraser](https://huggingface.co/datasets/NIATphysicalAI/eraser) | | 13 | `cube` | Stacking | 54 | 25,595 | 00:14:13 | 30 | fix required | [cube](https://huggingface.co/datasets/NIATphysicalAI/cube) | | 14 | `stapler` | Pick and place | 37 | 16,750 | 00:09:18 | 30 | OK | [stapler](https://huggingface.co/datasets/NIATphysicalAI/stapler) | | 15 | `bottle` | Orientation | 48 | 25,162 | 00:13:59 | 30 | OK | [bottle](https://huggingface.co/datasets/NIATphysicalAI/bottle) | | 16 | `penholder` | Sorting | 92 | 43,990 | 00:24:26 | 30 | OK | [penholder](https://huggingface.co/datasets/NIATphysicalAI/penholder) | | 17 | `sort_biodegradable` | Sorting | 169 | 87,635 | 00:48:41 | 30 | OK | [sort_biodegradable](https://huggingface.co/datasets/NIATphysicalAI/sort_biodegradable) | | 18 | `pick_place_cubes_para` | Pick and place | 300 | 92,287 | 00:51:16 | 30 | OK | [pick_place_cubes_para](https://huggingface.co/datasets/NIATphysicalAI/pick_place_cubes_para) | | 19 | `pick_place_cubes` | Pick and place | 300 | 92,287 | 00:51:16 | 30 | OK | [pick_place_cubes](https://huggingface.co/datasets/NIATphysicalAI/pick_place_cubes) | | 20 | `chocolate_pick_place` | Pick and place | 50 | 29,178 | 00:16:13 | 30 | OK | [chocolate_pick_place](https://huggingface.co/datasets/NIATphysicalAI/chocolate_pick_place) | | 21 | `pick_place` | Pick and place | 50 | 18,559 | 00:10:19 | 30 | OK | [pick-place](https://huggingface.co/datasets/Bradx86/pick-place) | | 22 | `push_t_v2` | Non-prehensile push | 194 | 79,027 | 00:43:54 | 30 | OK | [push_t_v2](https://huggingface.co/datasets/Bradx86/push_t_v2) | | 23 | `t_push` | Non-prehensile push | 50 | 9,388 | 00:05:13 | 30 | OK | [t_push](https://huggingface.co/datasets/Bradx86/t_push) | | 24 | `Bottle_orient` | Orientation | 50 | 32,196 | 00:17:53 | 30 | OK | [bottle-test](https://huggingface.co/datasets/Bradx86/bottle-test) | | 25 | `Cube_pick_place` | Pick and place | 50 | 28,764 | 00:15:59 | 30 | OK | [record-test](https://huggingface.co/datasets/Bradx86/record-test) | ## Repository Structure ``` NIAT10/ ├── / │ ├── meta/ │ │ ├── info.json # fps, robot type, feature schema, totals │ │ ├── episodes.jsonl # per-episode index and lengths │ │ ├── tasks.jsonl # natural-language task strings │ │ └── stats.json # per-feature normalisation statistics │ ├── data/ │ │ └── chunk-000/ │ │ └── episode_*.parquet # states, actions, timestamps │ └── videos/ │ └── chunk-000/ │ └── observation.images./episode_*.mp4 └── ... ``` ## Usage ### Authenticate ```bash hf auth login # or: export HF_TOKEN=your_token_here ``` ### Download a single dataset Each folder is self-contained, so pull only what you need: ```bash hf download NIATphysicalAI/NIAT10 \ --repo-type=dataset \ --include "Battery_sort/*" \ --local-dir ./NIAT10 ``` ### Download everything ```bash hf download NIATphysicalAI/NIAT10 --repo-type=dataset --local-dir ./NIAT10 ``` ### Load with LeRobot ```python from lerobot.datasets.lerobot_dataset import LeRobotDataset ds = LeRobotDataset(repo_id="NIATphysicalAI/NIAT10", root="./NIAT10/Battery_sort") print(f"Episodes: {ds.num_episodes}") print(f"Frames: {ds.num_frames}") print(f"Task: {ds.meta.tasks}") sample = ds[0] print(sample.keys()) ``` ### Browse the collection ```python from pathlib import Path import json for folder in sorted(Path("./NIAT10").iterdir()): info_path = folder / "meta" / "info.json" if info_path.exists(): info = json.loads(info_path.read_text()) hours = info["total_frames"] / info["fps"] / 3600 print(f"{folder.name:<26} {info['total_episodes']:>4} episodes {hours:.2f} h") ``` ### Load the tabular data without video Each folder is registered as a named config, so the state/action streams can be read directly with `datasets`: ```python from datasets import load_dataset ds = load_dataset("NIATphysicalAI/NIAT10", "Battery_sort", split="train") ``` ## Training ```bash lerobot-train \ --policy.type=act \ --dataset.repo_id=NIATphysicalAI/NIAT10 \ --dataset.root=./NIAT10/Battery_sort \ --output_dir=./outputs/act_battery_sort \ --batch_size=8 \ --steps=100000 ``` To train across several folders, pass a comma-separated list of roots or build a `MultiLeRobotDataset` over the folders you want. ## Known Issues and Caveats Community-recorded teleoperation data is never uniform. Things to check before training: - **Review status.** 2 of 25 datasets are flagged below. They are included for completeness; prefer the approved subset for headline results. - **Varying episode counts.** Some datasets are short (a few minutes) and will be heavily under-represented in a naive concatenation. Consider weighted sampling. - **Camera configurations differ** between datasets. Check `observation.images.*` in each `meta/info.json` before batching across folders. - **Frame rates.** 30 fps (25) — resample or filter if your policy assumes a fixed rate. - **Episode indices are per-folder.** They are not globally unique across the collection. - **Not a single merged dataset.** Loading NIAT10 as one `LeRobotDataset` will not work; point `root` at an individual folder. ### Flagged datasets | Folder | Status | Note | |---|---|---| | `Eraser_drawer` | fix required | Dataset card needs correcting | | `cube` | fix required | Frozen `camera2` feed on some episodes | ## Intended Use - Behaviour cloning and imitation learning on tabletop manipulation - Fine-tuning vision-language-action models on a consistent hardware setup - Multi-task and task-family transfer experiments - Benchmarking data-efficiency across task types - Teaching and coursework on robot learning pipelines ## Data Collection All episodes were collected by **human teleoperation** on an **SO-101 (leader–follower pair)** setup. An operator moved the leader arm by hand while the follower arm mirrored the motion, and joint states, actions and synchronised video were recorded through LeRobot. No scripted or autonomous policies were used, so every trajectory reflects human timing, hesitation and correction. ### Robot | Property | Value | |---|---| | Arm | SO-101 (leader–follower pair) | | Control | Leader arm (leader–follower joint mirroring) | | Recording framework | LeRobot | ### Cameras Two cameras per episode, both the same sensor: | View | Sensor | |---|---| | Overhead | IMX335 5MP USB Camera (B), 5V USB 2.0, 175° wide angle | | Wrist | IMX335 5MP USB Camera (B), 5V USB 2.0, 175° wide angle | The **overhead** camera gives a fixed third-person view of the whole workspace; the **wrist** camera is mounted on the follower arm and moves with the end effector, providing close-range detail during grasps and contact. The 175° field of view keeps the full table in frame from a short mounting distance, at the cost of noticeable barrel distortion near the edges — the recordings are **not undistorted**, so calibrate or rectify yourself if your method assumes a pinhole model. **Cropped overhead views.** Some datasets have a cropped overhead feed, done to remove distractors and give a bounded region of interest for training. Original 1920×1080 frames were cropped to a 955×720 region of interest (margins: 580 left, 385 right, 180 top and bottom) and downsampled to 640×480. The wrist view is uncropped. Because the crop changes the effective field of view and scale, a policy trained on cropped datasets will not transfer cleanly to uncropped ones without matching the preprocessing. Both streams were captured through LeRobot's OpenCV backend: ```json { "type": "opencv", "index_or_path": "/dev/video2", "width": 640, "height": 480, "fps": 30 } ``` Recorded at **640 × 480 @ 30 fps** — well below the sensor's 5MP capability, chosen to keep two USB streams stable and file sizes manageable. Device indices vary between recording sessions; check `observation.images.*` in each folder's `meta/info.json` for the exact keys and shapes a given dataset uses. ### Environment and randomisation In most datasets object positions were **randomised between runs**, so the policy cannot succeed by memorising a fixed layout. Randomisation was manual rather than programmatic, so coverage is uneven — some datasets vary position more aggressively than others. ## Language Variants Not every folder is an independent recording. One dataset re-annotates episodes that already exist elsewhere in the collection: | Variant | Source episodes | Paraphrases | Difference | |---|---|---|---| | `pick_place_cubes_para` | `pick_place_cubes` | 8 | Same episodes, task string expressed 8 different ways | `pick_place_cubes_para` contains the **same trajectories** as `pick_place_cubes`, with the natural-language task description rewritten in 8 different phrasings. It exists to test whether a language-conditioned policy generalises across instruction wording rather than latching onto one exact string. Two consequences worth knowing: - **Do not count both toward dataset size.** Together they represent one set of recordings, not two. The unique figures in the statistics table already exclude the variant. - **Do not put both in the same training mix without thinking.** Naively concatenating them duplicates every episode, doubling that task's weight. Either train on the paraphrased version alone, or sample the pair as one dataset. ## Limitations - **Environment.** Most recordings share one lab, one table and similar lighting. Background and surface diversity is limited. - **Single embodiment.** Everything is SO-101. Cross-embodiment transfer is untested. - **Fixed camera geometry** at 640 × 480, with uncorrected wide-angle distortion, and inconsistent overhead cropping between datasets. - **Uneven dataset sizes.** Durations range from a few minutes to nearly an hour, so naive concatenation heavily over-weights the longer datasets. - **Manual randomisation** means object-position coverage is not uniform or measured; in some datasets objects were re-randomised only every ~10 episodes. - **Human demonstrations only** — no failure cases, recovery behaviours or counterexamples, which limits use for methods that need negative data. ## License Released under the **Apache 2.0** license. Individual datasets may carry additional attribution requirements. ## Related Resources - [LeRobot](https://github.com/huggingface/lerobot) — framework used for recording - [LeRobot docs](https://huggingface.co/docs/lerobot) — full documentation - [Dataset format guide](https://huggingface.co/blog/lerobot-datasets) — best practices - [SmolVLA](https://huggingface.co/blog/smolvla) — VLA model these datasets suit - [Community Dataset v3](https://huggingface.co/datasets/HuggingFaceVLA/community_dataset_v3) — larger cross-embodiment collection