CRIT-IRL · Decision-Driven Phase Transitions in Biological Collectives via Interpretable Multi-Agent IRL with Closed-Loop VR and Robotic Platform Validation

Lines connect the coordinator with its partners.

Project objective

CRIT-IRL will uncover the individual decision rules that drive phase transitions in biological collectives and will validate these mechanisms causally using closed-loop virtual reality (VR) and bio-hybrid robotics. The project addresses a key gap: existing models reproduce global patterns but rarely capture interpretable, context-dependent decisions—or show how such decisions generate critical transitions at the group level.Objective 1. From trajectories to decision rules: infer interpretable, multi-objective policies from high-resolution multi-animal data via multi-agent inverse reinforcement learning (IRL) with graph-based interaction modelling; deliver a validated model, curated datasets and open code.Objective 2. From decision rules to criticality: construct phase diagrams linking decision parameters to collective regimes; derive early-warning diagnostics (variance, recovery time, influential-agent scores) and identify critical individuals/time windows precipitating transitions.Objective 3. Closed-loop causal validation: embed the learned policy in immersive VR with live fish and deploy it on aquatic robots to steer groups across phase boundaries, benchmarking against rule-based baselines and ablations.Methodological innovations. Interpretable MA-IRL with constrained reward bases and heterogeneity; multi-scale validation (micro kinematics to macro order parameters); physics-informed mechanism mapping; embodied, counterfactual tests in VR/robotics. Open science: preregistered analyses, FAIR data, and reusable protocols/software.CRIT-IRL yields mechanistic insight into decision-driven criticality, principled controllers for bio-hybrid swarms, and transferable diagnostics for other complex systems (e.g., ecological monitoring, distributed AI). The fellowship provides integrated training across behavioural biology, machine learning, complex-systems physics, and VR/robotics, leveraging state-of-the-art facilities to maximise scientific and career impact

Original text from CORDIS.

Participants

• MAX-PLANCK-GESELLSCHAFT ZUR FORDERUNG DER WISSENSCHAFTEN EV · MUNCHENCoordinatorGermany

Links

• View on CORDIS
• DOI: 10.3030/101278020