Research

Motion planning, robot learning, and autonomy under uncertainty.

My research and invention work sits at the intersection of planning, prediction, control, and machine learning. I separate it here into publications, which document peer-reviewed and project-based research, and patents, which document original autonomous-systems inventions.

Publications

DSCC 2019 Chance-constrained multi-agent planning with priority ordering, belief updates, and replanning. View section ACC submission / arXiv Trajectory-induced self-calibration for hidden-target localization through an unknown-pose relay. View section CDC submission / arXiv Excitation-supervised closed-loop self-calibration and target seeking. View section NeurIPS 2026 Global PIQA evaluates commonsense reasoning across 100+ languages and cultures. View section CTI 2020 CTI 2020 work on latch-state classification and Gaussian Process pressure modeling. View section

Featured Research

Receding horizon chance-constrained motion planning for uncertain multi-agent systems.

This project studies how autonomous agents can plan through cluttered environments while accounting for uncertainty in their own state, other agents, and moving obstacles. The implementation combines Chance-Constrained RRT tree expansion, priority-ordered trajectory prediction, Monte Carlo collision estimation, Kalman covariance updates, and receding-horizon replanning.

Explore Multi-agent CC-RRT Read the paper

Scalable Coordination Priority-ordered planning supports variable numbers of agents and moving obstacles.

Chance Constraints Monte Carlo collision checks bound risk under uncertain state estimates.

Belief Updates Kalman updates reduce covariance as new measurements arrive during execution.

Receding Horizon Lazy horizon checks can trigger replanning when predictions violate feasibility.

ACC submission / arXiv

Trajectory-induced self-calibration for hidden-target localization.

This paper studies the estimation problem: a vehicle receives local range-bearing packets from a relay whose global position and yaw are unknown, and uses its own motion to recover relay pose and the hidden target. The central result is that a single static pose leaves a continuous gauge ambiguity, while two distinct vehicle-relative observations are enough for constructive recovery in the noiseless one-relay case.

Explore adaptive localization Read ACC submission Code and data

Minimal Excitation Two distinct vehicle-relative observations can determine relay yaw, relay position, and anchored target in the noiseless case.

Conditioning Metric Trajectory spread connects the constructive recovery result to estimator conditioning and excitation design.

Noisy Recovery Monte Carlo studies report millimeter-scale target error under finite range-bearing noise and stress tests.

Robustness Sweeps The results study noise, initialization, outliers, dropout, pose error, and one-versus-two-relay behavior.

CDC submission / arXiv

Excitation-supervised closed-loop self-calibration and target seeking.

This paper adds the control layer around the self-calibration estimator. The robot monitors a trajectory-spread certificate online, keeps generating excitation when the stored packet window is weak, and transitions to unrestricted target seeking once the calibration certificate is strong enough.

Explore adaptive localization Read CDC submission Code and data

Spread Supervision The controller uses the trajectory-spread margin as a finite-noise calibration certificate.

Excitation Resets Exploratory motion is retriggered when the packet window is underexcited.

Paired Trials Monte Carlo comparisons quantify when supervision improves yaw accuracy and target-seeking success.

ROS 2/Gazebo Software-in-the-loop runs validate the loop with message transport, physics integration, and sensing delay.

NeurIPS 2026

Global PIQA: evaluating physical commonsense across 100+ languages and cultures.

Global PIQA is a multilingual and multicultural benchmark for evaluating physical commonsense reasoning in large language models. The dataset was constructed through a participatory effort by researchers across more than 65 countries, covering 141 language varieties, 19 language families, and 24 writing systems, with examples verified by native speakers.

Read the arXiv preprint Open PDF

Global Coverage 141 language varieties spanning five continents, 19 language families, and 24 writing systems.

Cultural Specificity Non-parallel examples emphasize local foods, customs, traditions, and everyday knowledge.

Native Verification Examples were checked by native speakers to improve linguistic and cultural validity.

LLM Evaluation The benchmark exposes large performance gaps, especially for lower-resource languages.

CTI 2020

Fast data-driven modeling of hydraulic clutch control pressure.

This paper studies pressure modeling for a hydraulic clutch control circuit with nonlinear behavior driven by hysteresis, latch transitions, solenoid dynamics, and ramp-rate effects. The modeling pipeline adds current-derivative features to capture direction-dependent pressure response, separates latch-related regimes through classification, and trains local Gaussian Process regressors for the resulting operating regions.

Read the arXiv preprint Open PDF

Regime Separation Latch and non-latch operating regions are separated before fitting pressure models.

Hysteresis Features Current-derivative inputs encode ramp direction and improve pressure-response modeling.

Local GP Models Gaussian Process regressors approximate nonlinear pressure behavior within classified regimes.

Amesim Comparison The learned model is compared against measured test-stand data and a physics-based baseline.