Humanoid Robotics Technology
AGIBOT has announced the open-source release of AGIBOT WORLD 2026, a heterogeneous dataset designed to support five key research pathways in embodied intelligence.
The dataset provides structured, precisely annotated real-world robot data, offering researchers and developers a robust foundation for training and evaluating next-generation embodied AI systems.
Expanding Data Collection Beyond Structured Demonstrations
AGIBOT WORLD 2026 captures data across a broad range of real-world environments, including commercial settings, homes, and everyday operational scenarios. The dataset is designed to reflect the variability and complexity that robots encounter in practical applications.
Unlike conventional datasets built on predefined and repetitive demonstrations, AGIBOT adopts a free-form data collection methodology. Teleoperators perform tasks dynamically in response to real-time conditions, enabling greater diversity across object types, task sequences, and environmental configurations.
To complement the real-world data, AGIBOT has also developed corresponding 1:1 digital twin environments, with simulation data released alongside the physical dataset to support sim-to-real research workflows.
Aligning Data with Real Robot Behavior
A central challenge in embodied AI is ensuring that datasets accurately represent how robots operate as integrated systems. To address this, AGIBOT WORLD 2026 incorporates several key capabilities:
- Whole-Body Control (WBC): Coordinated control of arms, waist, and hands to enable fluid, system-level task execution
- First-person beyond-visual-range teleoperation: Aligns operator perception with the robot’s viewpoint for more intuitive control
- Force-controlled data collection: Captures contact dynamics and physical interaction data alongside motion trajectories
These features are designed to ensure that the dataset reflects both motion and interaction, supporting more realistic and transferable model training.
Industrial-Grade Data Capture and Processing
Data for AGIBOT WORLD 2026 is collected using AGIBOT’s G2 hardware platform, which integrates high-performance joint actuators, multi-modal sensors, and a centralized domain controller for precise force control.
The system incorporates Zhixing 90D grippers and the dexterous OmniHand, enabling the capture of synchronized multi-modal data, including RGB(D), tactile signals, LiDAR point clouds, IMU data, and full-body joint states.
Each data episode undergoes validation and processing through an industrial-grade data pipeline to ensure consistency and readiness for large-scale training applications.
Phase One Focuses on Imitation Learning
AGIBOT WORLD 2026 will be released in five phases, each aligned with a specific research area in embodied intelligence.
The first phase focuses on imitation learning, providing hundreds of hours of real-world data collected primarily in commercial and service environments. The dataset includes:
- Task-level descriptions
- Step-by-step action sequences
- Atomic skill labels (e.g., pull, place)
- Object annotations, including bounding boxes and attributes
In addition, error-recovery trajectories are retained and annotated, enabling models to learn corrective behaviors and improve robustness.
Supporting Long-Term Development of Embodied AI
AGIBOT’s release of AGIBOT WORLD 2026 reflects a broader commitment to advancing the embodied AI ecosystem through open access to high-quality data.
By combining large-scale real-world datasets with corresponding simulation environments, the company aims to support global research efforts and accelerate the transition of embodied AI from experimental systems to real-world deployment.
Source: AGIBOT
