Embedded Files
Over the span of 3 undergrad years and 1 graduate year, I was first a state estimation engineer then a technical lead for MIT Driverless. We participated in the Indy Autonomous Challenge, as part of the greater MIT-Pitt-RW Team. We raced and had exhibition runs at the Indianapolis Motor Speedway, Lucas Oil Raceway Park, Texas Motor Speedway, Las Vegas Motor Speedway, Autodromo Nazionale Monza, Goodwood Festival of Speed, and the WeatherTech Raceway at Laguna Seca.
Some Appearances in Competition:
Las Vegas Motor Speedway, 2022
Texas Motor Speedway, 2022
Autodromo Nazionale Monza, 2023
Las Vegas Motor Speedway, 2024
During my time at MIT Driverless, I contributed towards the Perception and State Estimation autonomy stacks. At first, my code was deployed in the context of simulation races and time trial scenarios. As the Indy Autonomous Challenge reached maturity year after year, the competition evolved from single-car time trials to multi-car head-to-head racing. I then shifted to a Tech Lead for our rival estimation stack, focusing on reconstructing rival vehicle trajectories from imperfect data, and predicting rival long-term behavior and strategy beyond dynamical and racing line constraints.
Relying on onboard sensing only, what information can we assume about our rival's long term strategy, through a traceable prediction system built on fundamentals and learning?
Research Projects, in collaboration with MIT Driverless:
Multi-Agent Prediction for Autonomous Motorsports
Decoupled Multi-Modal Perception for Autonomous Racing
Robust Modeling & Controls for Racing on the Edge
Fast & Modular Autonomy Software for Autonomous Racing Vehicles
Multi-Agent Prediction for Autonomous Motorsports (2021-2023)
A course project for MIT's 16.412 - Cognitive Robotics (G). In this project, we use a Multi-Hypothesis Tracker to predict and accurately correct for the state vector of adversarial race cars, through exteroceptive detections from LiDAR, RADAR and Cameras. In our prediction setting, we predict the trajectory of the rival car up to 5 seconds in the future, beyond the domain that can be predicted by dynamic constraints alone. We build upon our Extended Kalman Filter backbone and integrate a learned model for intelligence and long-term behavior prediction. Learning the Hidden Markov Model that classifies rival car motion as overtaking, racing line, or blocking.
The learned Hidden Markov Model uses contextual detections in the scene to first apply the behavior prediction to a vehicle in a pedestrian-heavy and intersection setting, to estimate intention and future trajectory of other agents in the scene.
We use a Multiple Hypothesis Tracker to maintain several belief states of the rival car trajectories, and a scoring function that outputs the most likely hypothesis. Having multiple belief states allows us to remain accurate in the presence of occlusions, after overtaking, and as vehicles enter and exit the field of view. To prevent an unbounded growth of belief states, we prune the Tracker Hypothesis regularly.
This project included a presentation in the form of a 30 minute lecture about the topic and an extended abstract.
16_412_Grand_Challenge_Report.pdfLearned State Estimation Lab Meeting - Learned State Estimation Lab Meeting.pdf16.412 Lecture.pdfExtended Abstract
Excerpt from Internal Research Presentation
Topic Lecture Slides
2. Decoupled Multi-Modal Perception for Autonomous Racing (2022-2023)
An important consideration is that with estimates running at 10hz, at speeds of 200kph, our vehicle moves 5-6m per update, during which we are essentially blind. Our perception stack has to account for this latency and provide detections that prevent a scenario where a collision risk could happen during the blind period.
In this work, we describe our system for multi-modal perception to detect rival cars through three distinct and independent pipelines: Camera detections, LiDAR and RADAR. We then fuse these pipelines to obtain a single position estimate of the rival car. We use YOLO on our Camera pipelines and assumption-informed scene reconstruction to triangulate the position through the monocular or binocular setting. We use PointPillars on our LiDAR reconstruction to quickly identify the rival cars while ignoring other noise in the scene. We use relative movement of RADAR detections to identify rival vehicles.
We perform SLAM during our initial testing laps (track reconnaisance) and perform localization when racing at speed.
Our follow-on work, (1. Multi-Agent Prediction for Autonomous Motorsports) uses the fused detection pipeline to perform Multiple Vehicle Tracking across a Multi Hypothesis Tracker, with an enhanced state estimate.
Submitted for an ICRA 2022 talk. This work was further absorbed as the perception component and then accepted under our work Fast & Modular Autonomy Software for Autonomous Racing Vehicles, with the Journal of Field Robotics.
3. Robust Modeling & Controls for Racing on the Edge
ICRA Workshop
4. Fast & Modular Autonomy Software for Autonomous Racing Vehicles
Published with Journal of Field Robotics
More coming soon!
Page updated
Google Sites
Report abuse