Learned Search PruningQuadtree DecompositionMixture of Experts
Can learned guidance narrow a grid search at both coarse and fine map resolutions without giving up optimality? A small CNN over a quadtree decomposition predicts, level by level, which tiles can contain the optimal path, using one specialist network per map type and batched inference. Training uses teacher forcing followed by adversarial hard-example mining to focus on difficult cases, and a hybrid mode verifies the learned corridor with algebraic transfer matrices so the result stays optimal. Across five map families (random scatter, DFS maze, spiral, recursive division, rooms and corridors) the predicted corridor typically covers 8–15% of the grid while still containing the optimal path. Neural-only mode is fast but not guaranteed optimal; hybrid and matrix-only modes are. The mixture of experts needs the map type to pick a specialist, and a full runtime comparison against classical search is still open.
How can a robot reach a target in clutter when avoiding all contact is impractical? IMPACT considers both which objects tolerate contact and how pushing them may affect their surroundings. It combines VLM judgments of acceptable contact with sampled push outcomes to build a direction-aware costmap, then plans with a contact-aware A* that reasons about how pushes change the scene, planning through newly opened space while accounting for obstacles displaced into later parts of the route. My contribution focused on formulating the directional-contact representation and implementing the planner.
Could learned spectral response functions improve material classification beyond conventional three-channel imaging? This work developed a generative deep learning model, inspired by biological vision, that adapts spectral channels to the environment for more accurate and efficient material discrimination. Published patent application EP4339901A1.
Methane detection from hyperspectral satellite imagery as a one-shot problem: a contrastive model, trained with a triplet loss to learn spectral embeddings, compares an observed spectrum with a reference methane signature and predicts whether methane is present. Evaluated on AVIRIS data from the Aliso Canyon incident and published at IGARSS and WHISPERS 2023.
A 5x5 Go agent built for USC’s Foundations of Artificial Intelligence course, combining minimax search with alpha-beta pruning and a Q-learning component, with a tournament simulator and a small GUI for manual and automated play. Scored 95.2/100, in the top 5 of about 300 students.
