Programme: Horizon Europe cascade funding (FSTP), GRAIL project (grant agreement No. 101298421)
Objective of the call
GRAIL (Generative Robotics & AI for EU Industrial Leadership) is building a trustworthy, sovereign European foundation model for robotics. One of the biggest obstacles is the lack of high-quality data that robots need to learn useful tasks, adapt to changing conditions and recover when something goes wrong.
Genesis is the first of three GRAIL open calls. It funds projects that close this robotics data gap with high-quality multimodal datasets, affordable data-collection tools and reproducible benchmarks. Across the three calls, GRAIL plans to support at least 40 third-party projects.
Scope of activities
The call supports four types of activities:
- Generalist data foundations: multimodal datasets covering different robots, tasks and environments, such as grasping, pushing and object manipulation.
- Democratising data capture: affordable, easy-to-use data-collection tools such as sensor kits, mobile apps and gamified platforms.
- Lab-scale generalisation stress tests: tests that show how robots cope with changing conditions such as lighting, clutter and human presence.
- Open benchmarking tools: shared tests and evaluation methods to measure and compare robot performance in a consistent way.
Call topics
Each application addresses one of eight topics. The full scope, required data and deliverables for each topic are set out in Annex 1 of the Terms and Conditions.
- Dexterous and precise tasks with robotic hands: teleoperated demonstration data for multi-fingered hands with tactile sensing, including failure and recovery examples.
- Long-horizon stationary robot tasks: multimodal, multi-stage and contact-rich task data from real or realistic industrial settings.
- Long-horizon mobile robot tasks: data from warehouses, factories and logistics sites, focused on mobility and navigation.
- Industrial robot tool-use tasks: data of robot hands using real tools such as wrenches, hammers, screwdrivers and soldering stations.
- Egocentric video data collection for dexterous tasks: synchronised first-person video, wrist camera and force data of people performing dexterous tasks.
- Low-cost multimodal capture for contact-rich manipulation: plug-and-play, retrofittable sensing kits with a documented acquisition pipeline and a demonstration dataset.
- Plan–Fail–Repair, gamified collection of robot reasoning data: a game or gamified platform that records how robot plans are made, why they fail and how they are repaired.
- Open generalisation-stress benchmarking kit: physical stress artefacts, a sweep protocol and a portable evaluation harness with reliability and human-in-the-loop reporting.
Projects are expected to move from TRL 2 to TRL 4, although this is not mandatory. Integration with GRAIL models or platforms is only required where agreed in the mentoring plan.