This thesis develops autonomous robotic systems for data gathering in complex real-world environments. It advances three areas: informative path planning, disturbance-aware planning, and predictive world modelling. The research introduces adaptive sampling-based planners for long-horizon information gathering, methods for estimating wind fields and planning safe, energy-aware UAV trajectories, and predictive mapping systems for indoor exploration. It also compares robot exploration strategies with human decision-making and uses those insights to build a reinforcement learning planner that performs at near-human level. Overall, the work shows that integrating long-horizon reasoning, environmental dynamics, and richer belief models can make autonomous robotic information gathering more effective, reliable, and scalable.

In our complex world, how do humans learn and make decisions when their cognitive resources are limited? My thesis introduces a new theory called "policy compression" to answer this question! The basic idea is that people simplify their decision-making processes to reduce the mental effort required, without significantly compromising the benefits or rewards of those decisions. I use computational modeling, human experiments, and brain studies in rats to explain why people exhibit certain decision-making patterns, like the tendency to stick with familiar choices, and why they use strategies like "chunking" to reduce mental load. I also propose that different brain regions work together to balance mentally taxing decisions with more automatic, habitual decisions. This allows the brain to optimize behavior in complex environments. In conclusion, my thesis offers a new way to understand how humans and animals make decisions with limited mental resources, and shows how the brain organizes itself to handle decision-making efficiently.

This thesis proposes scaling humanoid robots through large-scale motion imitation, learned control priors, and reinforcement learning. A perpetual humanoid controller achieves full-dataset imitation, distilled into a universal latent space enabling efficient learning of manipulation and vision-based tasks. The approach transfers from simulation to real robots, advancing practical humanoid control.

My research develops navigable high-altitude stratospheric balloons that combine satellite-level coverage with drone-level detail at low cost. Using machine-learning trajectory models and altitude-based steering, fleets can monitor wildfires, deforestation, and environmental change in real time. This technology enables scalable, sustainable remote sensing for global environmental protection.