3D Perception
3D Scene Completion for Autonomous Driving
Developed a diffusion-based system that reconstructs complete 3D environments from sparse LiDAR scans. The pipeline was then distilled into a camera-only model — enabling a standard RGB camera to achieve spatial understanding comparable to a LiDAR sensor, with near-identical reconstruction quality. Developed using Claude Code for architecture iteration and evaluation tooling.
0.039 CD LiDAR model
0.040 CD Camera-only model
Under Review IEEE RA-L 2026
PyTorch Diffusion Models Point Clouds Cross-Modal Distillation
3D Perception
Sub-centimeter Camera Localization
High-precision camera tracking system for indoor environments using neural scene representations. Combines Monte Carlo particle filtering for pose refinement with CLIP-based retrieval for global initialization from an unknown starting position.
0.41cm Tracking accuracy
1.8cm Cold-start localization
3D Gaussian Splatting Monte Carlo Methods CLIP PyTorch
Robotics
Dual-Arm Robotic Teleoperation
End-to-end teleoperation pipeline for a dual-arm manipulation platform. Custom servo drivers, multi-camera depth integration, low-latency networking, and edge deployment on NVIDIA Jetson Orin for real-time onboard processing.
Dual-arm manipulation Real-time control
Jetson Orin Edge deployment
Multi-sensor Depth cameras + servo feedback
ROS 2 RealSense ZeroMQ Jetson Orin