The challenge has now come to an end. Thank you everyone for participating in the challenge and making it successful!
🏆 Check out the winners of the challenge here!
🛣. Real track scores are now updated
🚀 Fork the Starter Kit here
Simulation to real transfer is an important line of research for the development of autonomous vehicles. Generally this is a hurdle for individuals who want to try their ideas on a real car but cannot do so due to financial or practical limitations. The NeurIPS 2021 Deepracer challenge gives participants the unique opportunity to apply their deep learning expertise on a real world Deepracer car on a real world track. While it doesn’t emulate the full experience of the real world, it is a step in the right direction and an interesting space to benchmark on.
In this competition, you will train a reinforcement learning agent (i.e. an autonomous car), that learns to drive by interacting with its environment, a simulated track, by taking an action in a given state to maximize the expected reward. This model will then be tested on a real world track with a miniature AWS Deepracer car. Your goal is to train a model that can complete a lap as fast as possible without going off track, while avoiding crashing into the objects placed on the track.
The Deepracer challenge is a part of a series of embodied intelligence competitions in the field of autonomous vehicles, called The AI Driving Olympics (AI-DO). The overall objective of the AI-DO is to provide accessible mechanisms for benchmarking progress in autonomy applied to the task of autonomous driving. This edition of AI-DO will make use of the AWS DeepRacer platform. AWS DeepRacer is an AWS Machine Learning service for exploring reinforcement learning that is focused on autonomous racing.
🎮 The AWS Deepracer Platform
AWS DeepRacer is an integrated learning system for users of all levels to learn and explore reinforcement learning and to experiment and build autonomous driving applications. Follow the documentation link - What Is AWS DeepRacer?, to know more.
Deepracer Gym Environment - AWS Deepracer is a service hosted with AWS Robomaker platform. To make it easy for partcipants, we are releasing a gym environment for Deepracer.
🎯 Competition structure
The Deepracer challenge will have 2 rounds
Round 1 - Simulation Only
In Round 1, participants need to submit models that score higher on the deepracer simulator. The environment will behave like any other OpenAI gym environment and be subject to constraints on the time limit and the hardware used for the rollouts.
Round 2 - Sim2Real Transfer
In Round 2, models that perform above a certain threshold score will be tested on a real world track, the leaderboard will have scores for both the simulation and the real world tracks, with a higher importance given to the real world scores. More details will be announced when round 2 begins.
🎬 Starter kit
Get started with submissions using the starter kit. This will help you setup the environment easily.
You will be evaluated over several metrics, and each will be evaluated over multiple laps to reduce variance on the real world racetrack. The racetrack for the competition will be pre- specified but the obstacles will randomly placed on the racetrack. All the metrics will be eval- uated on the real-world racetrack as follows:
Number of laps: The number of times the AWS DeepRacer completes a full loop around the racetrack. The robocar may be go off-the-track up to 5 times before being disqualified. Every time the robocar goes off-the-track, the robocar will autonomously reset to the start-line on the racetrack.
Lap Time: The lap time will be measured as the fastest time around the racetrack after completing 3 laps.
Number of resets to the start line: To have a valid lap time no more than three resets will be allowed, otherwise the lap will be viewed as did not finish, and participants have to restart. Participant has to reset the car if all four wheels leave the track (which includes the border- lines)
Number of objects avoided: There will be a fixed number of randomly-placed objects on the race track during the evaluation. The number of avoided objects is the number of all objects minus the number of objects the robocar collided with.
Tracks on the evaluation server will have small variations from the tracks given out in the public release.
These metrics will be aggregated into a single score, which will be used for the leaderboard.
🚀 How to Make A Submission
Fork the Starter Kit here. It contains detailed descriptions of how to make your first submission to AIcrowd. If you have any questions regarding the submission process, ping us on Discord or post your question on the challenge forum. We'll help you resolve it at the earliest. 🙌
Round 1 (Simulation Round): September 6th -
October 31st November 10th
Round 2 (Sim2Real Round): November 17th - November 30th
Final Results Announced: 6th December 2021
🥶 Team Freeze Deadline:
October 31st November 25th
🏅 Top 10 Winners of Round 1 will receive AWS Credits worth $500 each
🥇🥈🥉 Top 3 Winners of Round 2 will receive AWS Credits worth $1,000 each
🥇🥈🥉 Top 3 Winners of Round 2 will all also receive 1 AWS DeepRacer Car each
Next 4 Participants on the Leaderboard of Round 2 will receive AWS Credits worth $500 each
📝 Participants with Top solutions will be offered Co-authorship for a research paper outlining the solutions, submitted to NeurIPS 2021 Competitions Track Proceedings.
💬 Discourse Forum
We strongly encourage you to use the public channels mentioned above for communications to the organizers.
Sahika Genc (AWS)
Dipam Chakraborty (AIcrowd)
Yoogattam Khandelwal (AIcrowd)
Siddhartha Laghuvarapu (AIcrowd)