What we do
We develop perception and autonomy for robots that operate in challenging, unstructured, and safety-critical environments.
Perception and autonomy for underwater vehicles operating in turbid, dynamic, and GPS-denied marine environments.
Multimodal 3D detection, reconstruction, and scene understanding from cameras, LiDAR, and sonar.
Fusing heterogeneous sensors and modalities into robust representations for real-world robotic decision-making.
Agents that perceive, reason, and act — closing the loop between perception, planning, and control on physical platforms.
Long-term autonomy for field robots, including navigation, mapping, and control in challenging environments.
Simulation-based data generation and digital twins that bridge the sim-to-real gap for perception models.
Efficient algorithms and hardware-aware frameworks that bring perception and reasoning onto resource-constrained edge devices.
Integrating large language models and knowledge graphs for higher-level robot reasoning and decision-making.
Selected work
A lightweight, near-identity module that stabilizes intermediate bird's-eye-view features of existing camera–LiDAR detectors, improving robustness under domain shift and sensor degradation without retraining the detection head.
A hierarchical VLM framework that decouples high-level semantic reasoning from low-level reactive control for autonomous underwater oyster reef monitoring.
A benchmark to measure common sense logical reasoning of Vision Language Models using curated synthetic images with defined logical anomalies accross real life categories to improve VLM preparedness for open world deployment.
A multi-institution effort on autonomous navigation and control for aquaculture monitoring, recognized with the best control framework award at the IROS workshop in Abu Dhabi. Led by Prof. Xiaomin Lin as lead and corresponding author.
Frameworks for autonomous underwater vehicles to detect and map marine objects such as oyster beds and coral reefs, combining real and synthetic data, simulation-based methods, and multimodal sensing.
Simulation-based synthetic data generation and digital-twin pipelines that bridge the sim-to-real gap for perception models deployed on field robots.
Efficient algorithms and frameworks for deploying perception and reasoning on edge devices, including integrating large language models with knowledge graphs for robot decision-making.
