Intelligence
in motion.

Vision-guided pick, inspect & sort

An in-development workcell concept for black automotive rubber bushings. The study uses a RealMan RM65-B 6-DoF arm and RH56F1 five-finger hand, with stereo depth for grasp planning. Additional close-up imaging is being evaluated to inspect surfaces that may be occluded during grasping.

Being evaluatedDiffusion-based grasp candidates, MoveIt 2 approach planning, and residual reinforcement learning for near-contact adjustments. These are planned modules, not verified production features.

Building an AMR from the electronics up

A differential-drive mobile robot capstone project (2026–present). Implemented work includes STM32–ODrive S1 CAN communication, drive sizing and torque analysis, power distribution and pre-charge PCB design, and chassis integration. ROS 2 navigation and multi-floor/elevator scenarios remain active development targets.

Camera, LiDAR, and RTK-GPS autonomy

A real 1/5-scale driving stack built with ROS 2: YOLO lane perception and bird's-eye-view geometry feed Pure Pursuit steering; 2D LiDAR supports Follow-the-Gap avoidance; RTK-GPS provides waypoint navigation. A control bridge arbitrates modes and implements timeout-based safe stopping.

Stanley tracking & Frenet obstacle avoidance

An F1TENTH/RoboRacer implementation integrating Stanley steering with a Frenet-frame local planner. The ROS 2 package generates nine lateral candidate paths, checks LiDAR and occupancy-grid constraints, and chooses a collision-free route around obstacles.

Multi-scale vision for end-to-end control

A CARLA behavior-cloning study comparing a ResNet-34/CILRS baseline against a multi-scale YOLO26s feature encoder for steering and longitudinal control. The proposed policy completed 5 of 5 evaluation routes in the internal simulation experiment; real-road performance is not implied.

Reported CARLA result5/5 route completions · approximately 2.5 km per run · test-only result