2026-challenge-demos
The 2026-challenge-demos dataset is a large-scale collection of 20,000 human-collected teleoperation trajectories designed for the 2026 BEHAVIOR Challenge. Maintained by the Stanford Vision and Learning Lab, this synthetic dataset is hosted on Hugging Face and provides 1,950 hours of high-quality demonstration data across 100 complex household tasks. It utilizes the LeRobot V3 format, incorporating RGB and depth observations, robot proprioception, and detailed skill annotations. The dataset is specifically engineered for training embodied AI agents in the OmniGibson simulation environment, focusing on long-horizon reasoning, navigation, and dexterous bimanual manipulation. With a total size of approximately 3.27 TB, it serves as a primary resource for researchers and developers participating in the CoRL 2026 challenge. The data is released under the MIT license, ensuring open access for academic and commercial use, and supports advanced imitation learning pipelines for house-scale robotic agents.
| Year | 2026 |
|---|---|
| Episodes | 20,000 |
| Total hours | 1,950 |
| Trajectories | 20,000 |
| Frames | 210,916,774 |
| Frame rate | 30 fps |
| Embodiments | JoyLo |
| Task categories | cleaning, dishwashing, folding, cooking, manipulation, pick-and-place, dexterous-manipulation, bimanual, long-horizon |
| License | MIT |
| Access | open — commercial use permitted |
| Maintainer | Stanford Vision and Learning Lab (OmniGibson) |
| Origin country | US |
The 2026-challenge-demos dataset represents a significant milestone in embodied AI research, providing a massive, standardized corpus for training agents to perform complex, multi-step household activities. Collected via the JoyLo whole-body teleoperation interface, the dataset captures the nuances of human interaction with articulated objects in the BEHAVIOR-1K environment. The collection methodology emphasizes high-fidelity physics and diverse task scenarios, ranging from simple pick-and-place operations to intricate cooking and cleaning sequences. Use cases include fine-tuning foundation models for robotics, developing robust imitation learning policies, and benchmarking agent performance in realistic, house-scale simulations. The dataset is structured to facilitate modular loading, allowing researchers to download specific task chunks or modalities as needed, thereby optimizing computational resources during the training process. By providing both raw HDF5 replay data and processed LeRobot-compatible files, the Stanford Vision and Learning Lab enables seamless integration with modern robotics research workflows.
Frequently asked questions
What is the primary purpose of the 2026-challenge-demos dataset?
It is designed to support the 2026 BEHAVIOR Challenge, providing training data for agents to solve 100 full-length household tasks in the OmniGibson simulator.
How is the dataset formatted?
The dataset follows the LeRobot V3 format, which is a widely-adopted standard for robot learning datasets, facilitating easy loading and processing.
Can I download only a portion of the dataset?
Yes, the dataset is stored in numbered chunks (e.g., chunk-000), allowing users to download specific tasks rather than the entire 3.27 TB repository.
What robot platform was used to collect these demonstrations?
The demonstrations were collected using JoyLo, a whole-body teleoperation interface that controls the robot base, torso, arms, and grippers.
Is the dataset available for commercial use?
Yes, the dataset is released under the MIT License, which permits both academic and commercial use.