AIRoA Mobile Manipulator Data Engine
The AIRoA Mobile Manipulator Data Engine is a massive real-robot dataset developed by the AI Robotic Association (AIRoA) to facilitate the training of Vision-Language-Action (VLA) foundation models. Centered on the Toyota Human Support Robot (HSR) platform, the engine provides approximately 10,000 hours of high-quality, multimodal data collected in realistic household environments. The dataset covers a wide spectrum of long-horizon tasks, including contact-rich manipulation like making coffee or hanging laundry, and complex navigation-manipulation sequences. Standardized in the LeRobot format, it integrates synchronized RGB-D streams, 6-axis force-torque signals, and hierarchical annotations. This initiative, supported by NEDO and released for the ICRA 2026 VLA Competition, is designed for researchers seeking to benchmark generalist agents across diverse, unstructured scenarios. Access is provided via a gated model on Hugging Face, ensuring a shared resource for the global robotics community to address safety, reliability, and sim-to-real transfer challenges.
| Year | 2025 |
|---|---|
| Total hours | 10,000 |
| Frame rate | 30 fps |
| Embodiments | Toyota Human Support Robot (HSR) |
| Task categories | manipulation, pick-and-place, cooking, cleaning, long-horizon, loco-manipulation |
| License | CC BY-NC-SA 4.0 |
| Access | gated |
| Maintainer | AI Robotic Association (AIRoA) |
| Origin country | JP |
Overview
The AIRoA Mobile Manipulator Data Engine (often referred to as the AIRoA MoMa Dataset in its initial release) is a large-scale project by the AI Robotic Association (AIRoA) aimed at bridging the data gap for general-purpose physical AI. Developed with support from Japan's METI and NEDO, the dataset represents one of the largest public efforts to document mobile manipulation in unstructured, real-world household environments.
Methodology & Collection
Data collection utilizes a fleet of Toyota Human Support Robots (HSR) operated via high-fidelity teleoperation systems (THSR). The collection process focuses on diverse household tasks across multiple domestic zones (kitchen, living room, bathroom). To ensure robustness, researchers implemented systematic domain randomization, including variations in lighting, object placement, and initial robot poses. A key technical feature is the synchronization of high-rate sensor streams, including 30Hz RGB-D video from head and wrist cameras, 11-DOF joint proprioception, and synchronized 6-axis wrist force-torque signals, which are critical for learning contact-rich interaction dynamics.
Task Hierarchy & Annotation
The dataset employs a unique two-layer hierarchical annotation scheme. Each long-horizon episode is annotated with high-level sub-goals (e.g., "open cabinet") and low-level primitive actions (e.g., "reach handle," "pull"). This structure facilitates research into hierarchical reinforcement learning and detailed failure analysis. The data is fully standardized in the LeRobot v2.1/v3.0 format, allowing for immediate integration with pre-existing VLA architectures like RT-1, OpenVLA, and SmolVLA.
Use Cases
- Foundation Model Pretraining: Using the 10,000-hour corpus to develop generalist VLA models.
- Sim-to-Real Benchmarking: Evaluating model performance in real-world scenarios through the ICRA 2026 competition.
- Contact-Rich Learning: Utilizing force-torque data to refine policies for delicate manipulation tasks.
- Error Recovery: Studying the included explicit failure cases to build agents capable of autonomous correction.
Frequently asked questions
What robot platform is used for data collection in the AIRoA Data Engine?
The primary platform is the Toyota Human Support Robot (HSR), which features a 4-DOF arm, 1-DOF gripper, 2-DOF head, and a 3-DOF omnidirectional mobile base.
What is the total scale of the dataset?
While the initial paper release covered approximately 94 hours (25,469 episodes), the full Data Engine released for the ICRA 2026 competition provides approximately 10,000 hours of robot data.
What data formats are supported by the AIRoA dataset?
The dataset is standardized using the Hugging Face LeRobot format (v2.1 and later), ensuring compatibility with popular robotic imitation learning libraries.
Does the dataset include force feedback information?
Yes, it includes synchronized 6-axis wrist force-torque signals (Fx, Fy, Fz, Mx, My, Mz) to support the learning of contact-rich tasks.
How can researchers access the full 10,000-hour dataset?
The dataset is hosted on Hugging Face under the 'airoa-org' organization as a gated repository; users must agree to the terms of use and share contact information for access.