Closed-loop waypoint navigation with AprilTag landmarks

UC San Diego, Sep 2026 – present (ongoing). Team of 2; responsible for scheme design and parameter tuning, calibration, sensor experiments, test tooling and real-robot validation. The navigation core was written by my teammate.

A small rover drifts when it executes motion primitives open-loop, so after every primitive it stops, observes AprilTag landmarks, updates an SE(2) extended Kalman filter with chi-square gating, and replans. A fine spin-calibration level cut the smallest executable turn from about 24° to 3.05° (CCW) and 4.55° (CW). Across 24 archived runs, the best run ended with estimated residuals of 3.1, 3.3 and 5.1 cm at the three waypoints (filter estimates, no external ground truth). The navigator does not depend on ROS and passes 93 offline test cases against a simulated rover with systematic error; the 24 runs logged 184 landmark observations, 42 of them rejected with the reason recorded.

Code on GitHub · Video

Probing music-theory concepts in a text-to-music diffusion model

UC San Diego, May – Jun 2026. Team of 4; responsible for the Mustango deployment, all Mustango probing experiments, and merging four models’ results into one notebook.

Does a latent-diffusion music generator encode music theory in its intermediate representations? Adapting the SynTheory probing method, I trained linear probes on Mustango’s VAE latents and UNet activations over a 17-run sweep of layers and diffusion timesteps. Chord-type accuracy reached 0.88 on 1,000 clips (chance 0.25), while pitch-class accuracy from VAE latents stayed near chance (about 0.10 against 0.08) even with frequency-preserving pooling. A handcrafted chroma control on the same split reached 0.72, so the bottleneck is in the latent representation, not in the pooling.

A reproducible GWAS pipeline: linear regression versus linear mixed model

UC San Diego, Jan – Mar 2026. Team of 3; independent technical support.

Population structure inflates association statistics, so this pipeline takes 1000 Genomes CHB chromosome 22, simulates a trait from 20 causal variants (seed 42), and runs PLINK linear regression against a GCTA linear mixed model, reporting λGC with Manhattan and Q-Q plots. An interactive driver checks the environment, runs the five steps in order, and retries or skips a failed step. Both methods gave λGC close to 1, a small change on a cohort that was already well mixed, so the run is a check of the pipeline rather than a demonstration of the mixed model’s advantage.

Code on GitHub