Simulation: Neural rendering, controllable diffusion models and reinforcement learning for photorealistic, safety-critical and diverse simulations.
Path prediction and planning: Leading solutions for diverse future trajectory prediction and LLM-assisted motion planning in complex road scenes.
2D perception: Open vocabulary object detection, instance segmentation and tracking to handle even rare, unseen categories.
3D perception: State-of-the-art in single-camera 3D pose using neural rendering and instantaneous motion from Lidar inputs.
Devops: Vision-language models and LLMs for automated training and verification, with continual learning and adaptation for rapid generalization.
Efficiency: Deployment of compact but accurate models, with multi-task inference and high performance with vision transformers.