Note

This is a Hugging Face dataset. For large datasets, ensure huggingface_hub>=1.1.3 to avoid rate limits. Learn more in the Hugging Face integration docs.

This is a subset of the agibot-world/AgiBotWorld2026 dataset created for use with FiftyOne.

Installation#

If you haven’t already, install FiftyOne:

pip install -U fiftyone

Usage#

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub("Voxel51/AgiBotWorld2026-sample-10ep")

session = fo.launch_app(dataset)

Dataset Card for AgiBot World 2026 (10-episode FiftyOne subset)#

AgiBot World 2026 preview

Dataset Details#

Dataset Description#

AgiBot World 2026 is a large-scale, real-world embodied intelligence dataset collected on the AgiBot G2 dual-arm humanoid platform. It spans three top-level splits on the Hugging Face repo — ImitationLearning, ReinforcementLearning, and RichInteraction (plus a separate simulation split from the companion GenieSim project) — covering commercial-space, home, and other general-purpose manipulation scenarios.

This subset contains 10 episodes taken from a single ReinforcementLearning shard (ReinforcementLearning/Home/task_12192/HG-DAgger/20857245_20857259.tar.gz) of the source repo, all for the task “Insert the loose change into the coin slot of the piggy bank dagger”. Each episode records the dual-arm robot performing this insertion task under HG-DAgger-style intervention/rollout collection, with 7 synchronized camera streams and full proprioceptive state/action vectors.

Curated by: AgiBot World Team (source data); FiftyOne subset prepared for browsing/exploration in FiftyOne.
Funded by: Not specified in the source dataset.
Language(s): English (task instruction text)
License: CC BY-NC-SA 4.0 (inherited from the source dataset)

Dataset Sources#

Repository: https://huggingface.co/datasets/agibot-world/AgiBotWorld2026

Uses#

Direct Use#

Exploring, filtering, and visually inspecting AgiBot World 2026 robot-manipulation episodes in the FiftyOne App; prototyping data loaders and visualization pipelines for LeRobot-format multimodal robot-learning data before scaling up to the full 12+ TB source dataset; understanding the shape and structure of AgiBot’s HG-DAgger reinforcement-learning collection format (reward/done/intervened/succeed/truncated per-frame signals).

Out-of-Scope Use#

Not intended as a representative sample of the full AgiBot World 2026 dataset — it covers a single task, single collection mode (HG-DAgger), and single category (ReinforcementLearning) out of the many tasks/scenes/categories in the source. Not suitable for training production policies given its small size (10 episodes). Commercial use is restricted by the CC BY-NC-SA 4.0 license.

Dataset Structure#

This dataset uses FiftyOne’s multimodal / LeRobot representation: one sample = one episode. Each sample references the episode’s per-frame Parquet data (state/action vectors) and its 7 per-camera MP4 video streams via a media_reference; there are no per-sample image/video fields directly on the sample.

Fields#

Field	Type	Description
`id`	`ObjectIdField`	FiftyOne sample ID
`media_reference`	`MediaReferenceField`	Pointer to the episode’s Parquet (state/action) and MP4 (camera) shard files
`episode_index`	`IntField`	Episode index within this subset (0-9)
`task`	`StringField`	Free-text task instruction for the episode (all episodes share the same instruction in this subset)
`tasks`	`ListField(StringField)`	List form of the task instruction(s) associated with the episode
`length`	`IntField`	Number of frames in the episode (2176-2582 in this subset)
`duration`	`FloatField`	Episode duration in seconds (72.5-86.1s in this subset)
`robot_type`	`StringField`	Robot platform, `g2a` (AgiBot G2) for all episodes
`fps`	`FloatField`	Frame rate, 30.0 for all episodes

task is a StringField rather than a Classification because the source stores it as free-text natural-language instructions, not a fixed label taxonomy.

Per-frame data is not materialized as FiftyOne fields — it is read on demand from the referenced Parquet/MP4 shards and surfaced in the FiftyOne App’s State & Action / Streams tabs. The per-frame schema (from meta/info.json) is:

Per-frame feature	dtype / shape	Description
`observation.images.top_head`	video, 400x640x3	Head-mounted top camera (AV1/yuv420p)
`observation.images.hand_left`	video, 528x640x3	Left hand/wrist camera (AV1/yuv420p)
`observation.images.hand_right`	video, 528x640x3	Right hand/wrist camera (AV1/yuv420p)
`observation.images.head_left_fisheye`	video, 768x960x3	Head left fisheye camera (AV1/yuv420p)
`observation.images.head_right_fisheye`	video, 768x960x3	Head right fisheye camera (AV1/yuv420p)
`observation.images.head_back_fisheye`	video, 768x960x3	Head rear-facing fisheye camera (AV1/yuv420p)
`observation.images.head_depth`	video, 400x640x3, depth map	Head depth stream, encoded as PNG-codec video with `gray16be` pixel format
`observation.state`	float32[216]	Full proprioceptive state vector: effector positions, end-effector wrench/velocity/orientation/position, joint position/effort/mode/velocity, head/waist position & mode, robot base position/orientation, plus 6 camera extrinsic rotation matrices + translation vectors (see `meta/info.json` -> `features.observation.state.field_descriptions` for the exact index layout)
`action`	float32[44]	Action vector: effector positions, end-effector position/orientation, joint position, head/waist position, robot base velocity (see `field_descriptions` for index layout)
`reward`	float32[1]	Per-frame RL reward signal
`done`	bool[1]	Episode-termination flag
`succeed`	bool[1]	Task-success flag
`intervened`	bool[1]	Whether a human intervened at this frame (HG-DAgger collection signal)
`truncated`	bool[1]	Episode-truncation flag
`action_prob`	float32[1]	Policy action probability/confidence at this frame
`episode_index`, `frame_index`, `index`, `task_index`, `timestamp`	int64 / float32	Standard LeRobot bookkeeping columns

Label types and why#

There are no detection/classification/segmentation label fields in this subset — it is raw multimodal sensorimotor data (video + proprioceptive state/action time series), not annotated with FiftyOne label types.

`dataset.info` contents#

{
    "lerobot": {
        "format": "LeRobotDataset",
        "format_major": 3,
        "episode_count": 10,
        "imported_episode_count": 10,
        "skipped_episodes": [],
    }
}

Parsing decisions#

Why these episodes: The source agibot-world/AgiBotWorld2026 repo (12.3 TB, 420 files) is not laid out as a standard flat LeRobot repo — it has no root-level meta/info.json. Instead it ships as per-task .tar.gz archives nested under ImitationLearning/, ReinforcementLearning/, RichInteraction/, and simulation/, each containing a self-contained LeRobot v2.1 export (data/, meta/, videos/) for one or more episodes of one task. This subset uses all 10 episodes from a single archive, ReinforcementLearning/Home/task_12192/HG-DAgger/20857245_20857259.tar.gz, chosen because it was the smallest available ReinforcementLearning archive (~3.8 GB) and contains a complete, internally consistent set of episodes for one task.
Why only ReinforcementLearning, not a mix of categories: ImitationLearning, ReinforcementLearning, and RichInteraction archives were sampled and inspected, but they declare incompatible per-frame schemas — observation.state is shape [193] for ImitationLearning, [216] for ReinforcementLearning (this subset), and [169] for RichInteraction, and only ReinforcementLearning episodes carry the reward/done/succeed/intervened/truncated/action_prob columns. A single LeRobot v3 export needs one consistent Parquet schema, so mixing categories was not possible; this subset stays within the single-schema ReinforcementLearning / HG-DAgger slice.
Excluded modalities: none dropped from the declared feature set. observation.images.head_depth is declared as dtype: "video" (not the FiftyOne-importer-unsupported depth_video), so it imported like an RGB stream even though it is a gray16be PNG-codec depth map — no pixel values are log/linearly re-encoded and no decode parameters are needed, unlike some other LeRobot depth conventions.
v2.1 -> v3.0 conversion: the source ships LeRobot v2.1. It was converted in place with the official lerobot.scripts.convert_dataset_v21_to_v30 (adding the total_chunks/total_videos/total_tasks fields the converter expects) before import. Extra top-level info.json keys not part of the LeRobot v2.1/v3 spec (camera_parameters, data_version, h5_path, instruction_segments, key_frame, take_over — AgiBot’s additional annotation layers for subtask segments, 2D bounding boxes, and instruction-level step labels) were dropped by the converter and are not present in this FiftyOne dataset or its re-export.
Re-export on push: FiftyOne’s push_to_hub re-exports the 10 selected episodes as a fresh, self-contained LeRobot v3 repo — episode_index is remapped to 0..9, task_index is remapped to the single task present, and per-episode/global statistics are recomputed from the 10 episodes actually included (they are not copied from the much larger source dataset’s statistics).
Codec caveats: the 6 RGB/fisheye camera streams use AV1 (libdav1d/yuv420p), which decodes fine in Chromium-based browsers; the depth stream uses a PNG codec with gray16be, an unusual combination that may not render as a normal video in all players/browsers.
Per-episode video statistics are zero: the source meta/episodes_stats.jsonl reports min/max/mean/std of 0.0 for every video feature — this is a known upstream artifact of AgiBot’s stats pipeline (documented on the source dataset card as not affecting usage), not an error introduced by this subset.

Dataset Creation#

Curation Rationale#

Created to make a small, browsable, schema-consistent slice of the very large (12.3 TB) AgiBot World 2026 dataset available for exploration in FiftyOne, and to document the dataset’s non-standard multi-archive repo layout and per-category schema differences for future subsetting work.

Source Data#

Data Collection and Processing#

Collected on the AgiBot G2 dual-arm robot platform in 100% real-world environments (commercial spaces, homes, and other general-purpose scenarios) via free-form teleoperation and, for the ReinforcementLearning category, an HG-DAgger-style human-in-the-loop intervention/rollout collection process. See the source dataset card for full collection methodology.

Who are the source data producers?#

The AgiBot World Team.

Annotations#

Annotation process#

No additional annotations were added for this subset. The source dataset’s extended annotation layers (subtask “Task Frame” segments, 2D bounding boxes, instruction-level “skill” segments, stored under key_frame/instruction_segments in the source meta/info.json) are not carried into this FiftyOne subset — see Parsing decisions.

Personal and Sensitive Information#

Episodes are recorded in real-world commercial and home environments and may contain incidental appearances of people, personal spaces, or identifiable objects/branding. No explicit PII redaction is documented by the source dataset.

Citation#

BibTeX:

@misc{agibotworld2026,
    title        = {AgiBot World 2026},
    author       = {AgiBot World Team},
    howpublished = {\url{https://huggingface.co/datasets/agibot-world/AgiBotWorld2026}},
    year         = {2026}
}

APA:

AgiBot World Team. (2026). AgiBot World 2026 [Data set]. Hugging Face. https://huggingface.co/datasets/agibot-world/AgiBotWorld2026

Dataset Card Authors#

Harpreet Sahota

Dataset Card Contact#