AgiBot World 2026 Dataset Targets Real-World Humanoid Training
AgiBot World has released the AgiBot World 2026 dataset, a large-scale resource designed to support research and development in embodied intelligence using humanoid robot platforms. The dataset, published on Hugging Face, is built entirely from real-world environments and collected using the company’s AGIBOT G2 robot.
2026 Humanoid Robot Market Report
160 pages of exclusive insight from global robotics experts – uncover funding trends, technology challenges, leading manufacturers, supply chain shifts, and surveys and forecasts on future humanoid applications.
Featuring insights from
Aaron Saunders, Former CTO of
Boston Dynamics,
now Google DeepMind
2026 Humanoid Robot Market Report
160 pages of exclusive insight from global robotics experts – uncover funding trends, technology challenges, leading manufacturers, supply chain shifts, and surveys and forecasts on future humanoid applications.
The dataset focuses on general-purpose scenarios across commercial and home settings, reflecting a shift toward training humanoid systems outside controlled laboratory conditions. According to the project description, all data originates from real-world interactions, with additional support from a simulation environment that mirrors these scenarios at a one-to-one scale through the GenieSim project.
At its core, the dataset is structured around episodic recordings of robot activity. Each episode includes synchronized state and action vectors, along with multi-camera video streams captured from head and hand perspectives. The data is stored in Apache Parquet format, with associated MP4 video files, enabling detailed analysis of both motion and perception during task execution.
A key feature of AgiBot World 2026 is its multi-layer annotation system. The dataset includes high-level task descriptions, subtask segmentation, object-level annotations using two-dimensional bounding boxes, and fine-grained instruction segments aligned with specific robot actions. These layers are designed to support a range of research directions, including hierarchical planning, imitation learning, and language-conditioned control.
The dataset follows the LeRobot v2.1 format, making it compatible with existing robotics training pipelines. It also provides tools to convert long-horizon episodes into smaller, single-instruction segments, which can simplify training workflows for policy learning and benchmarking.
From a systems perspective, the dataset captures detailed robot state information such as joint positions, velocities, and end-effector states, paired with corresponding action commands. This structure enables researchers to study closed-loop behavior and develop models that map perception to control in humanoid systems.
The release reflects a broader trend in humanoid robotics toward data-centric development. By combining real-world demonstrations with structured annotations and simulation counterparts, AgiBot World 2026 aims to bridge the gap between experimental research and deployable robotic capabilities.
The dataset is licensed under CC BY-NC-SA 4.0 and is available for download via Hugging Face, with partial samples provided for inspection. It requires Python 3.10 or later and PyTorch 2.2 or later when used with the LeRobot framework.
As humanoid robotics moves toward more general-purpose applications, datasets of this type are becoming foundational infrastructure. AgiBot World 2026 positions itself as a resource for training systems that can operate across diverse environments and tasks, with an emphasis on real-world fidelity and multi-level supervision.
Source: huggingface.co
