Fengcheng Yu

DriveReferee - Video Action Model Training and Evaluation. Toyota Research Institute collaboration. First author; Submitted to ICRA 2027

Research Assistant at USC (2026-04 – Present)

DriveReferee - Video Action Model Training and Evaluation. Toyota Research Institute collaboration. First author; Submitted to ICRA 2027

• Led Python/PyTorch training and evaluation for an 8B video-action generator conditioned on video history, ego state, and navigation; implemented reference-free pairwise flow-matching preference fine-tuning.
• Built geometric feedback from LQR tracking, kinematic bicycle rollouts, and footprint-boundary distances; used ground-truth maps to construct preference pairs and predicted maps to select trajectories at inference.
• Benchmarked ten methods on 12,146 NAVSIM scenes with six verdict sources and matched compute budgets; analyzed trajectory failures and measured FVD, LPIPS, and PSNR video quality.

Research Assistant at Sun Yat-sen University (2025-01 – 2025-06)

HouseTune - LLM Guided Conditional Diffusion. Co-author; AAAI 2026

• Contributed architecture and training/evaluation code combining LLM layout reasoning with conditional diffusion for physically constrained floorplan generation.
• Ran benchmark comparisons and ablations to assess generation quality and constraint satisfaction.

Research Assistant at Sun Yat-sen University (2024-03 – 2025-12)

OccLinker - Temporal 3D Occupancy Prediction. Co-first author; Pattern Recognition, Accepted

• Designed historical static-motion cross-attention for camera-based 3D occupancy prediction; owned PyTorch/MMDetection3D implementation, training, and evaluation across four backbones.
• Improved mIoU by up to 3.09 points and temporal consistency by up to 7.68 points; measured 10 ms added latency and 2.29% memory overhead.

Course Project at USC (2026-03 – 2026-05)

Edge FaaS - Runtime Optimization and Predictive Scaling. USC CSCI 599 course project

• Co-developed a C++/Python Linux serverless runtime using preloaded workers and copy-on-write, reducing worker startup from approximately 900 ms to 100 ms.
• Built automated burst-load trials and predictive scaling, reducing cold-start requests by 31% versus reactive scaling across five trials.

R&D Center Intern at Lingzong Technology (Shenzhen) Co., Ltd. (2024-07 – 2024-09)

• Modified multi-head SurroundOcc and SelfOcc camera-based perception models and compared spatial predictions across architecture variants.

Research Assistant at USC Networked Systems Laboratory (2025-10 – 2026-03)

CarGaze - Real-Time Multi-Camera Gaze Tracking. Co-author

• Integrated front/rear camera streams and geographic inputs; aligned timestamps within milliseconds and reduced multi-second processing lag through parallel buffering.

M.S. in Computer Science – University of Southern California (2027-05)

B.Eng. in Intelligence Science and Technology – Sun Yat-sen University (2025-06)