A brain-inspired cognitive architecture that makes AI agents dependable, autonomous and collaborative — demonstrated with a team of lunar robots coordinated by multi-agent reinforcement learning.
AI agents built on large language models are powerful but brittle: they can hallucinate, struggle to explain their decisions and are hard to trust in the physical world. ADA-COLLAB builds ADAa, a cognitive architecture inspired by Distributed Adaptive Control, in which large multimodal models become bounded components inside a structure with a world model, episodic memory, metacognition, theory of mind and an ethics-by-design engine aligned with the EU AI Act.
ADAa is demonstrated in three use cases: media and cultural heritage, stroke rehabilitation and space robotics. In the space use case, a heterogeneous lunar team — scout rovers, a manipulation rover and a lander — explores and samples terrain. The robots coordinate through communication-aware multi-agent reinforcement learning, run small language models on board and are supervised by a human operator through natural-language mission control. The project ends with a geological-sampling demonstration at DLR's lunar analogue facility.
SpaceR leads the work package on validation, use cases and demonstrators and the space use case. The team develops the requirements and benchmarking framework (building on the open-source Space Robotics Bench), natural-language mission control for rover fleets, adaptive guidance and control, and multi-agent reinforcement learning for decentralised coordination, tested in simulation, in the LunaLab and at DLR.
ADA-COLLAB has been selected for funding under Horizon Europe. Open positions are listed on the Join us page.