Mobile robots, including legged robots, humanoids, and drones, are increasingly deployed in unstructured and demanding environments—from scaling rubble and inspecting industrial facilities to navigating disaster zones.
These open-ended tasks require robustness to terrain, resilience under degraded sensing, and adaptability to diverse task complexity.
On the path to maturity in mobile robotics, data and benchmarks have proven to be the core driving factor, as demonstrated in autonomous driving.
There, large-scale datasets and task-specific benchmarks have propelled tasks such as localization, mapping and detection to reach remarkable levels of precision.
However, for mobile robots in the field, we are still missing this ecosystem. Unlike passenger vehicles, collecting data in the field with these platforms poses significant challenges.
The cost of collecting data is high, and these platforms cannot easily carry the full payload of sensors and ground-truth systems.
This raises several key open questions: What kind of datasets will have lasting impact and be most useful for robotics? Which sensors are truly essential to capture the task complexity of these platforms? What level of calibration and ground-truth effort is required to advance research in these tasks? Most importantly: how do we design benchmarks that are apt for the unique tasks and failure modes of these robots beyond navigation on suburban streets?
This workshop will bring together researchers from robotics, computer vision, and machine learning to define principles for the next generation of datasets and benchmarks.
By rethinking dataset and benchmark design in light of the distinct requirements of embodied intelligence, we aim to accelerate progress not only in localization, but also in perception, mapping, and long-term autonomy in the wild.
A full list of the other workshops and tutorials at IROS 2026 is available on the
official IROS 2026 Workshops & Tutorials page.
