Research Projects

3D Object Detection

We leverage fundamental computer vision principles and deep learning to advance automotive perception in the task of 3D object detection - the task of estimating the six degrees of freedom pose and dimensions of objects of interest.

3D Multi-Object Tracking

3D multi-object tracking is a key problem for autonomous vehicles, required to perform well-informed motion planning in dynamic environments. Particularly for densely occupied scenes, associating existing tracks to new detections remains challenging as existing systems tend to omit critical contextual information

Object-Level SLAM / 3D Reconstruction

Object-level pose estimation plays a key role in SLAM. The reliable object pose estimation can provide valuable information in many down-stream robotic applications (e.g., robot manipulation, navigation and autonomous driving).

Motion Prediction

Motion prediction is a crucial component in autonomous driving that enables self-driving vehicles to forecast where other road users (vehicles, pedestrians, cyclists) might be in the next few seconds, helping to ensure safe and reliable navigation.

End-to-End Autonomous Driving

End-to-end autonomous driving is a fully differentiable system that processes raw sensor data directly into driving actions or trajectories, eliminating traditional modular components in favor of a unified, jointly-optimized neural network

Open-World Embodied AI

Open-world embodied AI enables autonomous robots and vehicles to operate in diverse, unpredictable environments beyond pre-mapped domains. By leveraging multimodal sensor data and adaptive learning, these systems plan robust routes, handle unseen scenarios, and act without relying on dense prior maps or hand-crafted rules—advancing flexible, truly autonomous mobility.

Perception in Space

Computer vision excels in common, well represented domains. Perception in space is challenging due to largely being underrepresented in training datasets, which requires perception systems to efficiently use smaller sets of training data to successfully perform tasks in the domain.