AgiBot World 2026

AgiBot World 2026 (styled AGIBOT WORLD 2026 by its publisher) is a real-world robot manipulation dataset released in stages during 2026 by the Chinese robotics company AgiBot. It is hosted on Hugging Face as agibot-world/AgiBotWorld2026, uses the LeRobot v2.1 directory format, and is licensed under CC BY-NC-SA 4.0. The data was collected on the company's dual-arm AgiBot G2 platform in what the dataset card describes as "100% real-world environments" covering commercial spaces, homes, and other general-purpose scenarios (the Theme 3 release adds an Industry scenario folder).^[1][2]

The dataset is organized into themes that the company unlocks one at a time. The project site names three: Imitation Learning, Rich Interaction, and Reinforcement Learning. On September 4, 2026, AgiBot's US account on X announced that "AGIBOT WORLD 2026 Theme 3 is now open source: 11,430 trajectories across 14 real-world tasks, including successes, failures, and human corrections."^[5][6] The trajectory count is the company's figure; the Hugging Face repository lists the 14 tasks but does not publish a trajectory total.^[3]

AgiBot World 2026 follows the company's 2025 dataset, AgiBot World (also called AgiBot World Colosseo), which reported about one million teleoperated trajectories. The 2026 release is much smaller in raw count and differs in intent: instead of only successful demonstrations, it adds contact-rich edge cases, failed policy rollouts, and human interventions that the company says are meant for reinforcement learning, world model training, and failure-recovery research in embodied AI and robot learning.^[6][8]

Dataset facts

All figures below are taken from the publisher's dataset card, the Hugging Face API, and the company's own announcements; none has been independently audited.

Scroll sideways for more →

Themes and release timeline

The project site presents AgiBot World 2026 under a "Key Themes" heading with the note that "Subsequent themes will be gradually unlocked in 2026." A separate highlights section says the dataset "spans five major embodied AI research paths" and "will be released in five sequential phases." Only three themes are named on the site as of September 5, 2026; the site does not say what the remaining two will cover, and this article does not guess.^[6]

The site lists the first theme as Imitation Learning without a number in the fetched copy, while the second and third are explicitly labeled "Theme 2: Rich Interaction" and "Theme 3: Reinforcement Learning." The numbering used here follows the company's.^[6]

Scroll sideways for more →

Folder counts are from the Hub file listing on September 5, 2026 and can change as the company adds packages. The six simulation tasks are named in their folder paths: scoop popcorn to bucket, take bagged food to cart, take cup to cart, take drink to cart, take laundry detergent to cart, and tidy up food in freezer.^[2]

Theme 3: Reinforcement Learning

The Theme 3 README describes the release as "real-world robot trajectories and feedback fields for reinforcement learning research," combining "expert demonstrations, human-in-the-loop correction trajectories, and success/failure rollouts generated by deployed policies." Compared with the earlier themes it "adds force-related observations and reward-related fields that capture execution outcomes, policy deviations, and human interventions in real-world tasks."^[3]

Tasks

The README lists 14 tasks. Some carry two task IDs, which the README does not explain; the second ID in each pair (16655 to 16661) does not appear as a folder in the Hub listing as of September 5, 2026.^[2][3]

Scroll sideways for more →

The scenario assignment (Industry or Home) comes from the folder each task sits in on the Hub.^[2][3]

Trajectory types

Each task folder is meant to hold up to three subfolders, one per trajectory type. The README warns that "Not every task necessarily contains all trajectory types" and that "Some currently visible folders may still use temporary names while the final layout is being updated." The commit history shows this in practice: packages uploaded on August 4, 2026 went into folders named PolicyRollout, while the August 31 uploads used Rollout, the name the README documents.^[3][4]

Scroll sideways for more →

HG-DAgger refers to a human-in-the-loop variant of dataset aggregation in which a human takes over from a running policy rather than providing demonstrations from scratch; the README uses the name without further definition.^[3]

Labels and extra fields

Theme 3 ships "task-level and process-level annotation labels for task execution and policy feedback." The README also adds two force-related features for all trajectories and five fields for correction and rollout trajectories.^[3]

Scroll sideways for more →

Scroll sideways for more →

The project site says these annotations "can be further transformed into feedback signals required for reinforcement learning, supporting reward model and value model training, as well as research on task progress prediction, success detection, error recognition, risk warning, intervention timing prediction, and error recovery."^[6]

Collection platform and pipeline

The dataset card and project site describe the collection setup in the company's own terms. The G2 is presented as "a next-generation, industry-standard robotic platform" with "high-performance joint actuators, multi-type sensors, and a high-performance domain controller." The site says that, fitted with the Swift Picker gripper and the OmniHand dexterous hand, the platform "collects RGB(D), tactile, force, LiDAR point-cloud, IMU, and full-body joint-state data synchronously in a unified pipeline." Which of these streams appear in a given theme's files is defined per theme by meta/info.json; the root README's camera list, for example, contains RGB, depth, fisheye, and stereo camera keys but no tactile or LiDAR key.^[1][6]

Teleoperation is described as "BLOS first-person teleoperation," in which "the teleoperator connects to robots remotely over public networks and shares the perceptual space with the robot from a first-person perspective." Movement during collection uses "a flexible mobile chassis, head movements, waist rotation, and lift-pitch motions," and the site says that data collection "incorporates force control, including contact and force feedback information."^[6]

After collection, the company says data "is validated through policy training and real-robot testing" and then passes through "a rigorous DaaS industrial quality control pipeline, which involves multiple rounds of screening and cleaning against unified standards." DaaS is the company's data-as-a-service business, which the site also markets alongside the dataset. The 2026 highlights section says the task matrix "encompasses fine-grained and ultra-long-horizon manipulation, spatial navigation, dual-arm coordination, and multi-agent / human-robot collaboration," and that each episode "features the dynamic disturbances, spatial complexity, and interaction uncertainty of real environments." These are the company's characterizations; the Hub listing so far shows tabletop and shelf manipulation tasks in commercial, home, and industrial scenes.^[2][6]

A digital twin of each real scene is used for simulated collection, and the resulting simulation data is released alongside the real data through the company's GenieSim project. The simulation/ directory on the Hub is the visible result.^[1][2]

Data format

The dataset follows the LeRobot v2.1 layout and points users to the LeRobot installation guide pinned to the v0.3.3 tag, which requires Python 3.10 or later and PyTorch 2.2 or later.^[1]

Scroll sideways for more →

info.json is the main manifest. It carries dataset summary fields (robot_type, total_episodes, total_frames, fps, splits), the data_path and video_path templates, the features schema, and optional custom metadata (instruction_segments, key_frame, high_level_instruction, take_over, h5_path, camera_parameters). The README notes that per-episode video channel statistics in episodes_stats.jsonl "are set to 0 due to internal preprocessing."^[1]

State and action vectors

observation.state and action are flat vectors stored in Parquet. Their layout is described by a field_descriptions block in info.json, in which each sub-field carries a description, a dimensions count, optional per-dimension dims names, and the indices that locate it in the flattened vector. Example keys given in the README are state/left_effector/position, state/right_effector/position, state/joint/position, state/joint/velocity, action/joint/position, and action/robot/velocity. The README does not publish the total vector length; users must read info.json for each task.^[1]

Camera keys

Camera streams are observation.images.* entries, each describing video type, frame rate, codec, pixel format, and frame shape. The README lists nine common keys: top_head, hand_left, hand_right, head_depth, head_left_fisheye, head_right_fisheye, head_back_fisheye_color, head_stereo_left_color, and head_stereo_right_color. The directory example in the README shows only the first three, so the set present in any given task should be checked in its manifest.^[1]

A documentation mismatch

The root README's sparse-checkout example tells users to select 20260315/tar/task_3777 and 20260315/task_3777, and its sample-dataset note points to task3777/380098_380609.tar.gz. Neither path exists in the repository as listed on September 5, 2026: task 3777 lives at ImitationLearning/CommercialSpaces/task_3777/ and holds 24 packages with different frame-range names. The example appears to predate the reorganization into theme folders.^[1][2]

Annotation layers

The root README says meta/info.json "adds three annotation layers beyond the standard LeRobot spec," stored under the key_frame and instruction_segments keys and indexed by episode index as a string. The hierarchy runs from the episode's high-level label in tasks.jsonl down to step-level segments.^[1]

Scroll sideways for more →

The README describes Layer 2 as usable "to train object-conditioned policies, ground language instructions to visual regions, or study grasp selection," and Layer 3 as enabling "skill-level imitation learning, primitive discovery, and multi-granularity language conditioning." The Task Frame layer is also what the companion splitting tool uses to cut long episodes into single-instruction episodes.^[1]

key_frame entries come in two categories, single (single-frame) and dual (range). Typical fields are track, frame_type_name, start, end, frame_detail.comment, frame_detail.error_cause, frame_detail.restorable, and frame_detail.extra. The README lists six built-in frame type names: Error Frame, Success Frame, Intervention Frame, Single Frame, 2D Bounding Box, and Task Frame. The first three are how success, error, and intervention events are marked at the frame level; the Theme 3 README's label list (success, error, intervention, adversarial disturbance) maps onto this mechanism, though the root README does not name an adversarial-disturbance frame type.^[1][3]

The example instructions in the README are all long-horizon retail restocking: moving drinks and yogurt between a shopping cart and the shelves of a refrigerated cabinet. Four split-episode instructions quoted in the README differ only in wording ("third layer of the chilled cabinet," "third shelf of the chilled cabinet," "third shelf of the refrigerated case," "third layer of the chilled display case"); the README does not explain the variation.^[1]

Two usage modes and the split tool

The dataset card describes two ways to consume the data.^[1]

Scroll sideways for more →

The tool is distributed as split_episodes_tool.zip (7,181 bytes on the Hub). It requires Python 3.8 or later, FFmpeg on the PATH, and the pyarrow and numpy packages. Its command line is python split_episode.py <input_dataset_path> [--output_path <output>] [--num_workers <N>], with the output path defaulting to <input_path>_split and the worker count to 4. The README says the script copies the full dataset to the output path without modifying the original, slices Parquet files by frame range while resetting frame_index, index, timestamp, and task_index, extracts video segments with FFmpeg while preserving codec and pixel format, and rewrites meta/info.json, episodes.jsonl, tasks.jsonl, and episodes_stats.jsonl. In the README's worked example, two original episodes with one and four Task Frame segments become five single-instruction episodes.^[1][2]

Relation to AgiBot World 2025

AgiBot World, the 2025 dataset, was published with the paper "AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems" (arXiv 2503.06669, first author Qingwen Bu, submitted March 9, 2025 and last revised August 4, 2025). The paper reports 1,001,552 trajectories totaling 2,976.4 hours across 217 tasks, 87 skills, and 106 scenes, collected by more than 100 AgiBot G1 robots through teleoperation. The companion GitHub repository, OpenDriveLab/AgiBot-World, later labeled the full Beta release as 1,003,672 trajectories (about 43.8 TB) and the Alpha subset as 92,214 trajectories (about 8.5 TB), and describes the work as an IROS 2025 Best Paper Award finalist with a 2026 IEEE Transactions on Robotics version. The repository's own code is "built on LeRobot (dataset v2.1)," the same format version the 2026 dataset uses. GO-1, the vision-language-action model trained on the 2025 data, was open-sourced on September 19, 2025 according to the repository's news list.^[8][9]

Scroll sideways for more →

The 2025 repository README describes the goal as "paving the path towards the 'ImageNet Moment' for Embodied AI"; the 2026 card opens by saying the new release follows "AGIBOT WORLD's 'ImageNet moment'," a reference to ImageNet's role in computer vision that the company applies to embodied AI. The comparison between "1M+" and "11,430" is between a full-year dataset and a single theme of a staged release, so the two totals are not directly comparable, and the company has not published a cumulative 2026 count.^[1][5][9]

The agibot-world Hub organization also hosts AgiBotDigitalWorld, GenieSimAssets, GenieSim3.0-Dataset, EWMBench, and the AgiBotWorldChallenge-2025 and AgiBotWorldChallenge-2026 competition sets; the 2026 challenge data was last modified April 24, 2026.^[12]

License and access

The dataset card states that "All the data and code within this repo are licensed under CC BY-NC-SA 4.0," the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. That license permits copying and adaptation with attribution and share-alike obligations but excludes commercial use, so the data can be used for academic research and open experiments while commercial training would require a separate arrangement with AgiBot. The company markets paid data collection ("Request your OWN data with the same ENTERPRISE level quality") and a data collection kit on the same site.^[1][6][10]

Files are downloaded with git lfs and a sparse checkout, or by cloning the repository with GIT_LFS_SKIP_SMUDGE=1 to fetch pointers only. Individual task packages are large: the two packages in the cabinet-door task (6671), for example, are about 35.6 GB and 17.3 GB.^[1][2]

Coverage

Interesting Engineering covered the Theme 3 release on September 4, 2026 in a video item by Jijo Malayil, which was also syndicated on MSN. The piece repeats the 11,430-trajectory and 14-task figures and adds that "The dataset includes 1,024 successful policy rollouts and 1,369 failed rollouts." Those rollout counts do not appear in the dataset card, the Theme 3 README, or the X post, and they are attributed here to the outlet. The article also states that AgiBot "plans to expand the initiative with additional datasets, benchmarks, and research resources," consistent with the site's five-phase description.^[7]

AgiBot's own announcement was a two-minute video posted from the @AGIBOT_US account at 17:00 UTC on September 4, 2026, linking the project site and the Hub repository; the Interesting Engineering item carries a publication timestamp of 13:13 UTC the same day, before the post.^[5][7]

References

1. ^AgiBot World 2026 dataset card (README.md) - Hugging Face (AgiBot World Team), last modified September 1, 2026, accessed September 5, 2026.
2. ^Hugging Face API metadata and file tree for agibot-world/AgiBotWorld2026 - Hugging Face, accessed September 5, 2026.
3. ^AGIBOT WORLD 2026 Theme 3: Reinforcement Learning (README_THEME_3.md) - Hugging Face (AgiBot World Team), uploaded September 1, 2026.
4. ^Commit history for agibot-world/AgiBotWorld2026 - Hugging Face, accessed September 5, 2026.
5. ^AGIBOT WORLD 2026 Theme 3 announcement - X (@AGIBOT_US), September 4, 2026.
6. ^AGIBOT WORLD project site, "Key Themes" and "AGIBOT WORLD 2026 Dataset Highlights" - AgiBot, accessed September 5, 2026 (text read from the site's client-side bundle).
7. ^Video: Chinese firm releases 11,430 robot trajectories to advance research - Interesting Engineering (Jijo Malayil), September 4, 2026.
8. ^AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems - arXiv (Qingwen Bu et al.), March 9, 2025, revised August 4, 2025.
9. ^OpenDriveLab/AgiBot-World repository README - GitHub (OpenDriveLab and AgiBot), accessed September 5, 2026.
10. ^Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) deed - Creative Commons, accessed September 5, 2026.
11. ^Hugging Face API metadata for agibot-world/AgiBotWorld-Beta - Hugging Face, accessed September 5, 2026.
12. ^Datasets published by the agibot-world organization - Hugging Face API, accessed September 5, 2026.