Back to blog posts

How to easily (almost) win a NeurIPS AV workshop challenge

Wed Oct 11 202316 views

Yuan Tian, a recently graduated UBC CS MSc student, placed second in the NeurIPS 2022 Competition Driving SMARTS for both Track 1 and Track 2 in December 2022

Her thesis work, following this success, compared the performance of several commercially available social agent models against DRIVE, examining how well they can be used to train AV/ADAS reinforcement learning (RL) policies. Her study used an RL environment consisting of the SMARTS simulation platform populated with one of three different social agent models: DRIVE, SUMO, and ZOO to compare policies trained using each model respectively. Performance of these trained policies is evaluated on a list of test scenarios, including highway merging, T-intersection, 4-way intersection (with variable number of lanes) and so forth populated with NPC agents exclusively from either DRIVE, SUMO, or ZOO. 

After comparing each trained policy against each test scenario populated with agents controlled by each social agent model, it was found that policies trained in simulations populated with DRIVE agents performed best overall due to DRIVE agents being the most behaviourally-diverse and most realistically interactive. It was additionally found that DRIVE agents were not as over-cautious or under-cautious as other social agent models and exhibited realistic traffic density and diversity. 

Check out Yuan's thesis presentation in the video below for more details about her evaluation of DRIVE highlighting an example of how Inverted AI's products have been Driving Intelligence all along.