Robotics MS · UC San Diego

Nathan Van Utrecht

I work on state estimation, robot learning, and control. Right now I'm in Xiaolong Wang's lab, teaching simulators to behave more like the real robots they stand in for.

Before this: internships at Brain Corp, John Deere, and Grace Technologies, and two research positions at Iowa State.

Nathan Van Utrecht

Live: frontier exploration, A* planning, 2D lidar. Click to set a goal.

Estimation & SLAM
EKF, factor graphs (GTSAM), ICP scan matching, quaternion optimization
Robot learning
PPO, SAC, BC / GAIL / AIRL, sim-to-real, PyTorch, Gymnasium
Perception
YOLOv8, CNNs, vision transformers, segmentation
Planning & control
Kinematics, screw trajectories, feedforward + PI, Koopman models, ROS
Mechanical
Creo Parametric, Windchill, FEA, prototyping and validation
  1. Visual-inertial SLAM A joint EKF on SE(3) that fuses IMU kinematics with optical-flow stereo features, bounding the drift of IMU-only dead reckoning on a Clearpath Jackal. EKF, sensor fusion, stereo vision Estimation 2026
  2. LiDAR SLAM with factor graphs Odometry, ICP scan matching, and GTSAM pose-graph optimization. Proximity-based loop closures cut total graph error 65%, against 43% for fixed-interval ones. GTSAM, ICP, pose graphs Estimation 2026
  3. Mobile manipulator pick-and-place Kinematics, 8-segment screw trajectories, and a feedforward + PI controller for a KUKA youBot. End-effector error converges within 5 s with little to no overshoot. Kinematics, trajectory planning, CoppeliaSim Control 2026
  4. Koopman models with KANs Kolmogorov-Arnold Networks as learned Koopman observables for a soft robot arm. Needs under half the training data of a polynomial EDMD baseline and runs in ~2 ms. PyTorch, Koopman operators, soft robotics Learning 2025
  5. IMU orientation tracking & panoramas Projected gradient descent over unit quaternions tracks IMU orientation against VICON ground truth, then stitches camera frames into a 2048×1536 panorama. Quaternions, constrained optimization Estimation 2026
  6. F1Tenth autonomous racing A follow-the-gap planner on raw LiDAR that laps the Nürburgring in the F1Tenth simulator, with separate fast and smooth modes. ROS, LiDAR, reactive planning Control 2023
  7. PPO from scratch Proximal Policy Optimization written in PyTorch and trained on four Gymnasium tasks, from CartPole and LunarLander to MuJoCo Hopper and HalfCheetah. PyTorch, Gymnasium Learning 2024
  8. Basketball shot detection YOLOv8 and segmentation track the ball, and a fitted trajectory calls make or miss. YOLOv8, transfer learning Perception 2024
  9. Traffic signs under occlusion Five models compared on occluded signs. A small custom CNN hit 95.9% accuracy at 1 ms per image; ResNet-50 reached 98.2% but ran 4.6× slower. CNNs, ViTs, data augmentation Perception 2024
Date Where Notes
2026 – now Xiaolong Wang Lab Graduate researcher, UC San Diego Research

Using real-world robot data to make simulation more faithful, so learned policies survive the move to hardware.

TBD Brain Corp Engineering intern, San Diego Industry

Details coming soon.

2024 – 25 Coordinated Systems Lab Honors researcher, Iowa State Research

Stress-tested BC, GAIL, and AIRL under shifted physics and goals. Wrote it up as my honors thesis.

2024 TrAC REU Research intern, Iowa State Research

Benchmarked model-based and model-free RL for sim-to-real. SAC reached an expert policy 5× faster.

2023 John Deere Engineering intern, Product engineering Industry

Took a tool-storage bracket from Creo concept to physical test; FEA cut its weight 15%. Four CAD concepts for new tractor cab features.

2022 – 23 Grace Technologies Engineering intern, IIoT engineering Industry

Six Python validation suites (>80% less manual testing) and a field debugger that cut callbacks 40%.