Controller Synthesis with Neural Network Verification

We propose a novel Lyapunov ReLU network to synthesize controllers for nonlinear systems with stability guarantees. The network is designed with two key features:

Positive definiteness and a unique global minimum: Achieved by constructing the network from monotonic functions over half-spaces.
Nested level sets: Facilitates the maximization of the Region of Attraction (ROA) through MILP-based optimization.

The Lyapunov conditions are enforced by formulating a Mixed Integer Linear Program (MILP) during network training. This framework provides a rigorous method for stability analysis and control design, with ongoing work extending its application to more complex scenarios such as collision avoidance in polygonal environments.

Publications:

Lyapunov Neural Network with Region of Attraction Search

IEEE American Control Conference (ACC) 2024, Toronto, Canada.

Box-based Efficient Robot Mapping and Navigation

We propose BoxMap, a learning architecture for end-to-end high-level map construction that represents environments (e.g., rooms and doors) as topological graphs. Using a DETR-like framework, BoxMap predicts topological graphs from low-level measurements with key features, including box embeddings, which model rooms and their relationships using a CNN and an encoder-decoder transformer; a TSDF-based loss, enabling efficient, label-free training via the Truncated Signed Distance Function; and a hierarchical loss, which detects small features like doors by isolating them from larger objects such as rooms. BoxMap supports compact mapping by directly updating graph-based maps from laser scans to reduce memory use and enables efficient structural reasoning, inferring unobserved areas and achieving shorter exploration paths compared to traditional methods.

Publications:

BoxMap: Efficient Structural Mapping and Navigation

IEEE International Conference on Robotics and Automation (ICRA, under review) 2025.

Task-Driven Robot Navigation in Structured Environments

We trained a multi-task U-Net deep network to address the challenge of extracting high-level information from low-level 2D laser measurements for tasks such as mapping, decision-making, and trajectory planning. Leveraging a U-Net architecture with skip connections, our approach extracts semantic information from low-level inputs and uses multi-task learning to integrate geometric and semantic data into a shared latent space. This enables iterative trajectory planning to an exit. The framework accumulates occupancy and camera maps, detects exits at the start or upon reaching interim goals, selects goals from an exit heatmap, and plans trajectories on hallucinated occupancy maps. Topometric maps are generated with room nodes derived from centroid and SDF estimates, and edges determined by traversability between nodes. By integrating mapping, decision-making, and planning, this framework enables efficient navigation and exploration in complex environments.

[poster][video]

Publications:

Do More with Less: Single-Model, Multi-Goal Architectures for Resource-Constrained Robots

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2023, Detroit, USA.

Bearing-based Formation Control

Given multiple mobile robots with single integrator dynamics, we designed a bearing-based controller that can steer the system of robots from random positions to a specific shape while minimizing the trajectory length. We first simplified sufficient conditions on global convergence of a bearing-based formation controller. Based on that, we formulated an optimization problem to automatically tune the controller to minimize trajectory lengths. The resulting controller generalizes well to new initial conditions and topologies.

[poster][video]

Publications:

Bearing-Based Formation Control with Optimal Motion Trajectory

IEEE American control conference (ACC) 2022, Atlanta, USA.

Learning-based 2D Simultaneous Localization & Mapping

We used ResNet to simultaneously find relative pose (scan matching) and detect loop closure given two keyframe measurements, from multi-task training. After that, we applied Pose Graph Optimization to correct the pose estimation. To train the model, we collected data from ROS using RRT exploration package in multiple environments.

[Github]