Autonomous robotic systems nowadays operate in dynamic and uncertain environments, where challenges arise from partially known system-environment interaction models, sensing noise, stochastic system behavior, and varying operational goals and constraints. These challenges necessitate bringing about a high degree of intelligence to achieve robust and flexible autonomy with tremendous potential impact in defense, healthcare, and manufacturing. Such intelligence requires an effective integration of system dynamics modeling with operation planning and individual robot control.
In this talk, I will first demonstrate the benefit of an integrated approach using optical robotic manipulation of micro-scale objects as a case study. A high-fidelity Langevin dynamics simulator is developed to model the probability of trapping an object in an optical field. The model is used in a partially observable Markov decision process algorithm to plan paths for multiple objects concurrently with collision avoidance and recovery steps. Experiments show successful transport of 2 micron diameter silica particles across the imaged workspace leading to manipulation of biological cells indirectly using the particles as optical fingers.
A novel functional analysis-based regression algorithm is then presented to scale up such coordinated operations to large multi-agent systems by learning the optimal solutions of decomposed but similar planning problems that are modeled as stochastic integer linear programs. I will conclude by briefly discussing another successful application on human robot collaboration in assembly kitting, and outline future research directions.