I write real-time, testable C++ code for perception and planning features that are running on trucks on the road right now.

Perception:

• Wrote and integrated nonlinear dynamics models and filters to estimate the states of obstacles such as other vehicles and pedestrians. This improved tracking, reducing false motion and harsh brakes from false positive detections without increasing false negatives.

• Worked with the safety team to address test cases to improve obstacle detection.

Planning:

• Developed a feature to ingest noisy map data and estimate road grade and bank angles. This significantly improved the truck's performance when starting/stopping on hills, along with general improvements to normal driving.

• Developed a novel anytime search algorithm for the global planner. This enables nonconvex decision making while improving compute time.

• Integrated performance metrics to ensure features stayed within compute budgets, along with determining what parts of the planner needed optimizing.

Miscellaneous:

• Triaged issues reported by the test engineers. This involved visualizing data to determine whether the truck's behavior was expected and determine the cause if unexpected.

• Wrote a map visualization tool to assist in debugging behaviors and finding mapping errors.

• Helped onboard new employees.

Thesis: Efficient and Reconfigurable Approximate Value Functions for Task Scheduling, Path Planning, and Control.

Multi-Robot Systems Lab

Out of Distribution Detection for Image-Based Systems Modeled in the Latent Space

• Developed a method for OOD detection to use when planning in the latent space of a learned model

using an autoencoder self-consistency metric.

• Works alongside previous work on graph-based methods.

Path Planning and Control using Learned Dynamic System Models

• Worked on tree- and graph-based algorithms to build value functions for and to control systems

whose dynamics are treated as a black box. This allows the methods to be applied to learned models

of the systems without needing physics equations.

Task Scheduling

• This work focused on the persistent surveillance problem, which requires drones to monitor a region

while intelligently charging, enabling long-term coverage. The problem has direct applications to

autonomous taxis and warehouse management.

• Developed an algorithm that reduced the problem to a tractable system to achieve near-optimal

performance, giving greater than 10x time improvement over optimal methods while significantly

outperforming heuristic-based methods.

AA 212 Advanced Feedback Control Design