Anderson Compalas

acompalas@ucsd.edu

I am a first year MS student in the Intelligent Systems: Robotics and Control program at UC San Diego. I received a B.S. in Applied Mathematics at UC Santa Cruz in 2025. As an undergraduate at UCSC, I had the opportunity to work with Eric Wang, tutoring for his class in machine learning, and helping out with research in multi-modal AI at his lab.

Right now I am looking to pursue research in deep reinforcement learning and Vision-Language-Action models, with applications in imitation learning, dexterous manipulation, legged robotics, and humanoids.

CV  |   GitHub

Dynamic programming solution to the Door & Key grid-world navigation problem, formulated as a deterministic MDP. Computes optimal policies for 7 known maps and a single unified policy for all 36 random 10x10 environments via state augmentation.

Visual-inertial SLAM using an EKF on SE(3), fusing IMU dead reckoning with stereo camera landmark observations to jointly estimate robot trajectory and a sparse 3D map. Includes extra-credit feature detection and tracking on dataset02 via Shi-Tomasi corners and Lucas-Kanade optical flow.

LiDAR Semantic Segmentation on the Waymo Open Dataset

Comparison of semantic segmentation architectures on the Waymo Open Dataset LiDAR subset across 23 classes, including a 2D U-Net on range images, PointNet, PointNet++, EdgeConv, an Eigen Feature baseline, and a diffusion-based approach. EdgeConv achieved the best result at 30.12% mIoU.

Proof-of-concept study replacing DP3's per-point MLP encoder with a frozen Point-JEPA self-supervised transformer, evaluated on three Adroit dexterous manipulation tasks. JEPA-64 exceeds the DP3 baseline on the door task; results on hammer and pen reveal the limits of global SSL pretraining for fine-grained visuomotor control.

Full LiDAR SLAM pipeline on a differential-drive robot combining encoder/IMU odometry, ICP scan matching, probabilistic occupancy mapping, RGBD texture mapping, and GTSAM pose graph optimization with loop closure.

Projected gradient descent on quaternion manifolds to track 3D orientation from IMU data, with panoramic image stitching from a rotating camera. Validated against VICON ground truth on 9 training datasets.

Built a ROS2-controlled kid's electric Jeep with closed-loop steering via AS5600 encoders, skid steering, and a complete ROS2 infrastructure including serial Arduino bridge and hardware E-stop.

Autonomous lap driving on the DonkeyCar platform using deep learning, GNSS, and OpenCV evaluated in both simulation and on hardware.

Developed simulation environments and diverse terrain conditions for quadruped locomotion on the Unitree B1 using NVIDIA Isaac Sim/Lab, supporting reinforcement learning-based locomotion research at the HARE Lab.

A full-stack fixed-wing UAV flight simulator built in Python, featuring 6-DOF aircraft dynamics, aerodynamic force and moment modeling, Dryden wind gust simulation, PID-based autopilot control, realistic sensor noise models, and complementary-filter state estimation.

website template