Safe and Agile Autonomy

A car trained to follow the 'rules of the road’ will perform well most of the time, but it is the unusual conditions, the edge cases, that pose the hardest safety challenges.

If everything seems under control then you are not going fast enough F1 racing

In our research we are teaching autonomous cars to learn how to drive at the limits of the control, and agility of the vehicle. The way we do that is by autonomously racing these cars against each other, both in highly photorealistic simulation and then on the 1/10 scale. Racing and driving fast, and driving safely may seem as two very contradictory objectives, but the idea is not to drive fast all times, but enhance the autonomous vehicle with the ability to be able to brake aggressively and maneuver aggressively, when it encounters a safety critical situation. We want to turn the unexpected situations and edge cases, to hedge cases, which you can bet your lives on. In formula 1 racing and motorsports in general, there is a saying that if everything seems under control, then you are not going fast enough. In a racing setting, we counter unusual situations, wheel to wheel action, and edge cases more often than regular driving. We are developing algorithms for structured deep neural networks that can learn an agile vehicle controller from expert driver behavior by watching annotated videos of how racing drivers, drive at the limits of control. We aim to demonstrate that a vehicle equipped with this agile controller leads to increased overall safety.

Research

Deepracing AI research encompasses high fidelity simulations to train the autonomous vehicle to be agile, to 1/10 scale fully autnomous race car testbeds, to testing on real work data and vehicles.

Learning to race using Formula 1 games

Our end-to-end deep neural netowrk learns to drive autonomously inside a photorealistic gaming environement.

F1/10 Autonomous Racing Testbed

Using a mix of manual (FPV) driven cars and autonomous cars, we can create annotated data-sets for edge cases which require agile control and agressive maneuvering.

1/10 Scale Testing

The DeepRacing AI learns to autonomously overtake using end-to-end learning. We test it on our F1/10 autonomous racing car platform.

Game to Track

Learning to race on the Melbourne Grand Prix Track in Australia in the F1 game and testing on real world F1 footage at the same track.

Safety benchmarking

Using customized environements in AirSim, we can create edge cases and evaluate if an agile autonomuos vehicles is safer.

Our Team

Meet our autonomous racing drivers

Trent Weiss

PhD, Computer Science

Dipshil Agrawal

MS, Computer Science

Varundev Sureshbabu

PhD, Computer Engineering

Jingyun Ning

MS, Computer Engineering

Shrirag Kodoor

MS, Computer Science

Dr. Madhur Behl

Assistant Professor, Computer Science, University of Virginia

Drop us a line

Our address

UVA Link Lab
151 Engineer's Way,
Charlottesville, VA 22902,
United States

  • Email:madhur.behl@virginia.edu
  • Phone:+1 434 924 1021

We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.