Algorithms · scientific computing · embodied AI
I work on efficient algorithms for structured computational problems, and on robotic systems that connect perception with action.
Semiring-based sparse algebraic path algorithms unify shortest paths, widest paths, and reachability, with open-source Julia implementations for CPUs and NVIDIA GPUs.
PATHPACK.jl is the accompanying software for this manuscript. Its documentation includes algorithms, examples, and the CPU and GPU interfaces.
A silent recording of the robot exploring and building an indoor map.
A Fetch robot builds a geometric map once, then updates an open-vocabulary semantic object graph across revisits as objects move, appear, or disappear.
My contributions include the RGB-D open-vocabulary perception pipeline and semantic object association and update logic.
* Equal contribution
A real-world perception benchmark across scene changes: 100 scenes, 70 objects, and 133,121 frames, with 49 models tested in 65 model–task evaluations.
Six tasks cover image depth, video depth, object tracking, relative camera pose, general VQA, and in-context VQA. The dataset includes over 5.2 million VQA question records.
Figures, findings, and demonstration
A U-Net predicts local coverage gain from partial maps, guiding memory- and structure-aware planning for efficient indoor robot exploration.
The public implementation connects coverage-gain heatmaps to candidate exploration directions, junction-aware planning, and frontier fallbacks.
A layered approach to multi-agent coordination using cellular sheaves and constrained harmonic extension. CellularSheaves.jl provides companion software.
Questions and implementations under active investigation.
A robotics course project exploring arm-driven mobility on a passive omni-wheel platform, with simulation in NVIDIA Isaac Sim.
A graduate coursework report, kept separately from the research publication list.