Nelson Tian

End-to-end interactive perception for unseen objects — no CAD needed. One click reconstructs a metric-scale 3D model and tracks 6D pose in real time (color–shape particle scale recovery + FoundationPose/ICP), reaching ADD(-S) 0.95 and 96% RSR (baselines: 0) at 19 FPS.

Independently developed the vision system for the Cloud Robotics track of the 11th RGMC (IEEE ICRA 2026, on KTH's CloudGripper platform): fisheye-undistorted Otsu segmentation + 360° template alignment producing a real-time IoU signal at ~6 ms/frame. The vision system helped the team place 2nd worldwide.

Voice grasping as a referring-expression task: Whisper → YOLO-Seg masks → optical-flow screening → Qwen-7B VLM grounding → Atom SDK grasp. Runs the VLM on the robot's onboard RTX 5000 within 9 GB VRAM; merged into Dobot's official repo.

Full autonomous navigation stack on a physical Limo robot: GMapping SLAM for environment mapping, AMCL localization, Pure Pursuit path following, and real-time YOLO-based traffic light detection — all integrated in ROS.

Engineered Fractal ArUco marker detection from scratch for the OpenCV objdetect module. Optimized the core algorithm structure and patched existing vulnerabilities, boosting keypoint matching capability by 28%. Contributed via PR #27934.

Evaluated 1,200+ samples to systematically benchmark ArUco3 detection performance. Optimized default parameters of the ArUco3 detector, improving robustness across lighting and scale conditions. Contributed via PR #28645, merged into OpenCV mainline.

Novel action-perturbation saliency method using output probability distribution gradients (measured by Euclidean distance) to explain agent decisions in Procedural Content Generation. Accelerated saliency computation by 10× compared to naive baselines.