Research

Classical MAPF solvers assume perfect centralized control, but real robot deployments face partial observability, communication latency, and human-in-the-loop needs. HI-MAPF introduces a tiered conflict-repair architecture: robots resolve local collisions decentrally while a lightweight human interface handles higher-level replanning when needed.

We quantify coordination overhead via a novel Information Units (IU) metric and validate the full system on TurtleBot4 hardware — one of the first MAPF works with an end-to-end physical robot deployment study across benchmark environments.

Controlling how opinions propagate through large social networks is inherently a planning problem — given a network state, which nodes to intervene on and in what sequence? We model this as generalized planning over graph-structured belief states, enabling a single policy to transfer across network topologies without retraining.

The InfoSpread framework trains Graph Convolutional Networks on PDDL-derived plan traces and packages the result as an interactive tool: upload any social network, configure spread parameters, and visualize optimal intervention strategies step by step.

Milestones include AAAI 2024 Demo, NeurIPS 2024, and AAAI-GenPlan 2025.

Automated planning lacks a shared vocabulary for communicating why a plan was chosen. We built PO (Planning Ontology) — an OWL 2 ontology covering planners, plan structures, domain models, and PROV-O provenance — so agents can annotate plans with SPARQL-queryable explanations. A web-based PO tool lets users explore competency questions interactively.

maPO extends PO to multi-agent settings, capturing agent roles, shared resource conflicts, and inter-agent dependencies. The OMEGA platform demonstrates this live: given any MAPF execution, OMEGA generates structured natural-language explanations grounded in the maPO ontology.