GrandTour — Legged Robot Dataset in the Wild
GrandTour is a large-scale multi-modal legged robot dataset collected by the Robotic Systems Lab at ETH Zurich using an ANYbotics ANYmal-D quadruped robot equipped with the Boxi open-source sensor payload. Released in 2026, it is the largest publicly available legged robot dataset to date, spanning 49 diverse indoor and outdoor environments including alpine terrain, forests, industrial sites, urban areas, university campuses, and train stations under varied conditions including day and night, rain, smoke, water, and sand. The dataset includes data from three LiDARs, ten cameras, one RGB-D camera, six depth cameras, eight IMUs, twelve proprioception sensors, and dual RTK-GPS, with gold-standard ground truth from a Leica Geosystem totalstation and 12,000 USD IMU. ANYmal-D walked 50,000 steps across 49 environments generating 150,000 images and 40,000 LiDAR point clouds. Access requires free registration via the project website. The dataset is designed for legged robot localisation, perception, and state estimation research.
| Year | 2026 |
|---|---|
| Embodiments | ANYbotics ANYmal-D |
| Modalities | rgb, lidar, depth, imu, proprioception |
| View type | multi-view |
| Task categories | locomotion, outdoor-navigation, inspection |
| Data format | rosbag |
| License | MIT — confirm in repository LICENSE file before activation |
| Access | gated — commercial use permitted |
| Maintainer | Robotic Systems Lab, ETH Zurich |
| Origin country | CH |
What is it?
GrandTour is a large-scale multi-modal legged robot dataset collected by the Robotic Systems Lab at ETH Zurich using an ANYbotics ANYmal-D quadruped robot equipped with the Boxi open-source sensor payload. Released in 2026, it is the largest publicly available legged robot dataset, spanning 49 diverse indoor and outdoor environments including alpine terrain, forests, industrial sites, urban areas, university campuses, and train stations under varied lighting and weather conditions including day and night, rain, smoke, water, and sand. The sensor suite includes three LiDARs, ten cameras, one RGB-D camera, six depth cameras, eight IMUs, twelve proprioception sensors, and dual RTK-GPS. Ground truth is provided by a Leica Geosystem total station and a high-grade IMU.
Who is it for?
GrandTour is designed for researchers working on legged robot localisation, state estimation, terrain-aware navigation, and multi-modal perception. It is particularly valuable for teams developing mapping, SLAM, and path planning systems for quadruped robots operating in unstructured real-world environments. The breadth of environments — from laboratory-grade indoor spaces to challenging alpine outdoor terrain — makes it useful for testing generalisation of perception and navigation systems.
Key specifications
- Robot platform: ANYbotics ANYmal-D quadruped
- Environments: 49 diverse indoor and outdoor locations
- Sensor suite: 3 LiDARs, 10 cameras, 1 RGB-D camera, 6 depth cameras, 8 IMUs, 12 proprioception sensors, dual RTK-GPS
- Ground truth: Leica Geosystem total station, high-grade IMU
- Steps recorded: 50,000+ walking steps
- Images: 150,000+
- LiDAR scans: 40,000+ point clouds
- Format: ROS bag
- License: MIT (confirm in repository LICENSE file before use)
- Access: Gated — free registration required at project website
How it compares
GrandTour is the largest publicly available legged robot dataset by environmental diversity and sensor modality count. Previous legged robot datasets such as HILL and RUGD cover outdoor terrain but with fewer sensor modalities and less environmental variety. GrandTour's 49-environment span and comprehensive ground truth make it the reference dataset for legged robot localisation research. However, it focuses on perception and navigation rather than manipulation or task learning, making it complementary to arm-focused datasets like EgoVerse.
Limitations and access notes
GrandTour covers locomotion and perception only — it does not include manipulation, object interaction, or task demonstrations. The ANYmal-D platform is a quadruped, not a humanoid; the dataset is not directly applicable to bipedal robot training. Access requires free registration via the project website. The MIT license covers the dataset itself; confirm the LICENSE file in the repository before commercial use.
Frequently asked questions
What robot is the GrandTour dataset collected on?
GrandTour is collected using the ANYbotics ANYmal-D quadruped robot, a legged robot designed for industrial inspection and unstructured terrain navigation. It is equipped with the Boxi open-source sensor payload developed at ETH Zurich's Robotic Systems Lab.
How do I access the GrandTour dataset?
Access requires free registration at the GrandTour project website (grand-tour.leggedrobotics.com). The dataset is gated — you must apply and receive approval before downloading. There is no direct Hugging Face or GitHub download link.
What environments does GrandTour cover?
GrandTour covers 49 environments including alpine terrain, forests, industrial facilities, urban streetscapes, university campuses, underground tunnels, and train stations. Conditions include day and night lighting, rain, smoke, water, and sandy surfaces.
Can GrandTour be used commercially?
The dataset is intended to be licensed under MIT, which permits commercial use. Confirm the license in the repository LICENSE file before use, as the license status was pending confirmation at time of indexing.
Is GrandTour useful for humanoid robot research?
GrandTour is collected on a quadruped robot (ANYmal-D), not a humanoid. It is not directly applicable to bipedal locomotion training. However, the sensor data and terrain diversity are valuable for perception and localisation systems that could transfer to humanoid platforms.