Multi-Robot Intelligence

A 4-year research program that grew from a single question — how do cooperative robots build a shared semantic understanding of the world? The answer became a hierarchical probabilistic framework for collaborative semantic map fusion, which then branched into cross-robot relative localization, aerial-ground heterogeneous mapping, and multi-vehicle formation control.

The theoretical foundation. The key contribution is the mathematical modeling of the hierarchical semantic map fusion framework and its probability decomposition. At the single-robot level, information from heterogeneous sensors is fused into a semantic point cloud, then converted into a local probabilistic semantic map via Bayesian label and occupancy updates. At the multi-robot level, since voxel correspondences between local maps are unknown, an Expectation-Maximization approach estimates the hidden data associations, and Bayesian rule merges semantic and occupancy probabilities into a globally consistent map.

A prerequisite for map fusion: determining the relative pose between robots that share no common reference frame. Existing approaches rely on low-level geometry features (points, lines, planes), which fail when initial error is large or overlap is low. COSEM jointly performs multimodal information fusion, semantic data association, and optimization in a unified framework. Instead of hard geometry data association, semantic and geometry associations are jointly estimated via Expectation-Maximization. The result is a rigid transformation matrix that aligns two semantic maps.

Collaborative heterogeneous robots face a fundamental challenge: UAVs and UGVs differ hugely in perception, mobility, and processing capabilities. This work formulates the UAV-UGV collaborative mapping problem with a continuous-discrete model. The key contribution is an information-gain trigger mechanism that determines when each robot should project its continuous local estimates into discrete space for multi-robot map fusion — solving the critical question of when to fuse, not just how. In discrete space, a probabilistic fusion algorithm based on Bayesian rule addresses multi-resolution hybrid map fusion between aerial and ground observations. Validated in GNSS-denied scenarios including tropical rainforest and multi-level buildings.

Extends collaborative perception into coordinated multi-vehicle action. The problem: UGV platoon cruise control in environments without positioning infrastructure — no lane markings, no roadside units, no GNSS. An integrated localization and planning framework composed of three modular algorithms: collaborative localization through matching local perceptions of different vehicles, trajectory reconstruction of the preceding vehicle to plan the follower's target state, and a virtual-controller that generates feasible trajectories in real time. The framework does not depend on direct observations between vehicles, making it applicable in non-line-of-sight situations.