Feb. 2023 - Present
Ideated and developed a motion planning system to guardrail ML model outputs and provide guarantees against egregious behaviors in autonomous driving using objective-driven AI methodologies. Delivered the software framework from scratch and guaranteed against 5+ bad driving behavior in over 8 software releases. This system served as the safety layer in behavior planning and played a pivotal role in enabling driverless vehicle operations.
Enhanced the machine learning model for trajectory selection by improving and developing new input features, model tuning, and data optimization to improve driving behavior and meet the company’s safety bar for autonomous driving.
Pioneered the estimation of epistemic uncertainty of ML models in motion planning. Developed a measure to quantify the correlation between data scarcity and output uncertainty. This enabled targeted data mining to improve data coverage for multiple driving modalities.
February 2023 - March 2024
Developed robust motion planning solutions for safety around pedestrians by enhancing constraint generation and cost design to improve trajectory generation and selection, solving 36% of failures seen on the road.
Collaborated cross-functionally to solve high-frequency degraded state failures by improving code quality, unit test coverage, and hardening planning algorithms which reduced vehicle retrieval events by 10x.
Revamped the dataset generation pipeline for the ML models used in AV behavior planning to enable caching and optimally running modules resulting in 5x cost savings and 20% pipeline runtime reduction per dataset generation job.
June 2022 - September 2022
Expanded the capability of the trajectory planner of the self-driving car by adding aggressive swerving capabilities leading to a 13% improvement in safety metrics in on-road scenario simulations.
July 2020 - Aug. 2021
Led the development of the localization and path planning stack for a drone-based autonomous inventory management product, called FlytWare, resulting in 5 pilot deployments across USA and Europe.
Built and deployed intelligent return-to-home path planner and a fail-safe state machine for end-to-end autonomous drone operations platform, FlytNow, which were monetized as value added services for FlytNow Enterprise.
Collaborated with senior engineers to design and implement the architecture of drone docking station module for FlytNow that could seamlessly integrate with docking stations from 5+ industry-leading manufacturers.
July 2020 - Aug. 2021
Built a C++ interface between the drone’s flight stack and Microsoft Airsim to introduce the capability of more realistic simulations that improve SITL testing by increasing recall in 3 identified risk areas, and improving scenario selection.