AgiBot World 2026

AgiBot World 2026 is an open-source heterogeneous robot learning dataset released by AgiBot in April 2026, collected entirely in 100% real-world environments across commercial spaces, homes, and general-purpose scenarios. It is collected on the AgiBot G2 hardware platform using a free-form collection mode, where teleoperators dynamically design task steps based on real-time conditions rather than following scripted demonstrations. The G2 — equipped with Swift Picker and the dexterous OmniHand — captures synchronized multimodal data including RGB(D), tactile signals, force, LiDAR point clouds, IMU, and full-body joint states in a unified pipeline. The dataset is released in five phases, each aligned with a core embodied-intelligence research direction; the first phase focuses on imitation learning. Digital twin simulation data is concurrently open-sourced via the GenieSim project. It succeeds AgiBot's earlier AgiBot World Alpha/Beta datasets.

Dataset specifications
Year2026
EmbodimentsAgiBot G2
Modalitiesrgb, depth, tactile, lidar, imu, proprioception, language
Task categoriesmanipulation, pick-and-place, long-horizon, bimanual, outdoor-navigation, human-robot-interaction
Data formatparquet, mp4, json
LicenseCC BY-NC-SA 4.0
Accessgated
MaintainerAgiBot
Origin countryCN

What is it?

AgiBot World 2026 is an open-source heterogeneous robot learning dataset released by AgiBot in April 2026, designed to systematically support five key research pathways in embodied intelligence. Unlike conventional datasets built on repetitive scripted demonstrations, AgiBot World 2026 uses a free-form collection strategy where teleoperators dynamically perform tasks based on real-time conditions, increasing per-episode diversity across object categories, initial configurations, and task sequences. The dataset is collected 100% in real-world environments — commercial spaces, homes, and general-purpose scenarios.

Who is it for?

Researchers working on imitation learning, whole-body humanoid control, and sim-to-real transfer, particularly those targeting wheeled humanoid platforms with dexterous hands. The rich multimodal sensor suite — including tactile and force data alongside RGB(D), LiDAR, and IMU — makes it valuable for contact-aware manipulation research where most datasets provide vision and proprioception only.

Key specifications

How it compares

AgiBot World 2026 succeeds AgiBot's earlier AgiBot World Alpha (595 hours, 36 tasks) and Beta (1M+ trajectories, 2,976 hours) datasets. The 2026 release differs in its free-form collection methodology, the newer G2 platform, and the addition of force-controlled data collection capturing contact dynamics. Compared to RoboMIND (107,000 trajectories, Apache 2.0), AgiBot World 2026 offers richer sensor modalities including tactile and LiDAR but under a more restrictive non-commercial license.

Limitations and access notes

CC BY-NC-SA 4.0 prohibits commercial use and requires share-alike on derivatives. Total episode and hour counts are being released in phases and are not finalized as a single figure. Chinese-origin dataset — geopolitical flag applies for IP awareness. Access requires Hugging Face login and condition acceptance.

Linked professions

Frequently asked questions

What robot is AgiBot World 2026 collected on?

AgiBot World 2026 is collected on the AgiBot G2, a wheeled humanoid platform with high-performance joint actuators, articulated head and waist movement, and lift-pitch capability. It is equipped with the Swift Picker and the dexterous OmniHand end effectors.

Can AgiBot World 2026 be used commercially?

No. It is licensed under CC BY-NC-SA 4.0, which restricts use to non-commercial research and requires derivative datasets to share under the same license.

What sensor modalities does AgiBot World 2026 include?

The G2 captures synchronized RGB(D), tactile signals, force, LiDAR point clouds, IMU data, and full-body joint states in a unified pipeline. The inclusion of tactile and force data is notable, since most large manipulation datasets provide only vision and proprioception.

How does AgiBot World 2026 differ from AgiBot World Alpha and Beta?

The earlier Alpha (595 hours, 36 tasks) and Beta (1M+ trajectories, 2,976 hours) datasets used a fixed task set. AgiBot World 2026 introduces free-form teleoperation for greater diversity, uses the newer G2 platform, and adds force-controlled data collection capturing contact dynamics. It is released in five phases.

What is free-form data collection?

Rather than following predetermined scripts, teleoperators dynamically design task steps based on real-time conditions in the environment. AgiBot states this improves generalization across object categories, initial configurations, and task execution sequences by increasing diversity within each episode.