Hi, I'mXilin Zhu

Giving a collaborative robot a sense not only of what to do but of when to move — by injecting the collaborator's 3D hand position and velocity into a VLA policy as a low-dimensional state token, and testing whether that reduces premature action without shortening the execution horizon.

A first-person AI twin whose only source of truth is material I reviewed and approved myself — with its safety and cost boundaries written as hard gates that refuse to start the service rather than as warnings.

A full ROS 2 driving stack on the KAL research car — RGB-D perception, overtaking and cone-corridor planning, Pure Pursuit plus PID control. I mainly owned perception; the detector trained on 400 real-car images reached 97.4% mAP50.

Built VDA 5050 fleet management and motion control over MQTT, then moved from simulation to two real mobile robots of different vehicle types, with type-aware path planning and station locking, verified in a live final assessment.

Improved YOLOv5s and C3D for static and dynamic gesture recognition (81.3% mAP, up to 83.3% accuracy), then used static gestures to drive a Raspberry Pi robot car over Wi-Fi in real time with 10 motion commands.

PyTorch implementations of a generative model and the core components of a Transformer vision architecture, covering the full path from building to training to inference.