Designing Intelligent Agents for Real-Time Experimental Control and Multi-Task Generalization

Scientific discovery has traditionally relied on human-led iterative loops of observation, modeling, and intervention. This thesis explores the possibility of automating components of this loop using artificial intelligence (AI), particularly in systems characterized by non-equilibrium dynamics, high dimensionality, and emergent behaviors. Two foundational challenges are addressed: automating physical modeling and enabling adaptive interaction with dynamic experimental systems, and generalizing agent behavior across tasks and contexts without retraining. To address the first challenge, we introduce a hierarchical AI framework for controlling active biomolecular matter, exemplified by microtubule–kinesin networks driven by light-activated motors. At the foundation are predictive models that learn the system’s response to static light patterns, enabling inverse design by selecting inputs that yield desired structural outcomes. Building on this, dynamic models construct low-dimensional representations of the system’s evolving state under time-varying stimuli, supporting forward simulation and real-time tracking. At the highest level, reinforcement learning agents—trained in simulation—discover and execute closed-loop control policies that achieve fine-grained manipulation objectives. These agents are deployed across ~100 parallel experimental setups, demonstrating autonomous operation with robustness, scalability, and reliable transfer. To address the second challenge, we investigate how generalist reinforcement learning agents can be constructed by leveraging the geometry of policy parameter space. We show that agents trained on distinct tasks self-organize into functionally segregated regions of weight space that encode both task identity and strategic variability. This insight enables the design of a hypernetwork—a network that generates the weights of other networks—that can interpolate smoothly between tasks and strategies via a single scalar input. Combined with a meta-controller, this architecture enables real-time modulation of agent behavior—ranging from conservative to risk-seeking—without retraining. Together, these contributions demonstrate that intelligent systems can both design and control physical experiments in real time, and adapt cognitive strategies across tasks through principled representations in policy space. This work establishes a foundation for closed-loop scientific autonomy, programmable biomaterials, and generalist AI agents, converging at the intersection of machine learning, biophysics, and automation.