As the AWS DeepRacer League finalists are gathering and finding their ways around Las Vegas, the AWS DeepSanta has come to town to bring good news: we’ve all been on the nice list!
Last week Jeff Barr announced new features on the AWS blog, he also gave an awesome shout out to us, the AWS DeepRacer Community. Throughout the week we could see changes to underlying packages and we’re tracking what we should be expecting.
Overnight the changes described went live in the console. Have a look around and let us know what you think.
Just a quick recap: the three main big changes in the console are:
- The Garage where you can set up your car with different sensors and network setup,
- ability to run your own tournaments
- training for three types of objectives: time trial, obstacle avoidance or head-to-head racing
The Garage
The Garage gives you a couple options including sensors, network topology, action space and updating your car’s look. It’s unclear to me if that is just for the rendered graphic or anything more (interestingly, the shell shown is for single camera regardless of the sensors that you select) :
One of the noticeable changes in action space is speed range which now has a maximum value of 4 m/s. This is a fix to the problem from season 2019 where values were at a limit of 12 m/s but in reality they were approximately 3 times smaller when you looked at your car’s speed in the simulation.
More technical note for those interested: This information is stored in model_metadata.json file which has an updated structure. It now contains version information (if missing it’s considered version 1, currently if present it’s set to 2), network topology and sensors listing. I have been assured by the AWS DeepRacer Team that old models are still supported, but bear in mind new ones will not work on your car until you update the firmware. Depending on how old your firmware is the process may be more or less complex. Refer to the documentation if you find any issues or ask the community.
New types of training
I reset my resources just to be on the safe side – since there were so many updates, my settings might be stale. While I was waiting for them to be recreated, I checked out the newly updated developer documentation about selecting the sensors.
In training you now get to choose your goal – time trial, object avoidance or head-to-head racing, and a car that you want to use for the training. For head-to-head you get to choose a setup for the bots that will take part in the training:
The rest of the setup looks fairly normal – set up the reward (there are new examples), choose the algorithm from a set of one, tune the hyperparameters, select training duration, start.
20191203-h2h-dr
Check out this video on Streamable using your phone, tablet or desktop.
Private races
I have created a race to test the experience. The steps on how to organise your race are listed in the documentation. In current version AWS covers the cost of submissions evaluations just like in the virtual league.
The set up was reasonably easy: you set up a starting and ending date, track (only tracks available in the console are available at the moment), name, description and a graphic to display. You also set up the rules: You get a choice of 3-7 lap attempts and 1-3 consecutive laps required to be listed.
I chose the wide reInvent 2018 track and name wiDR-2020 (not sure why 2020, but the first part is pronounced “wider”). It gives three tries per submission and all three need to be completed to get listed.
In the setup a joining link is created so that whoever receives it can join your race. I submitted one of my models and secured a win.
Racer’s experience in the community race is same as in the official virtual league in previous years. I am expecting that with 24 virtual league races and community races a lot will be happening throughout the 2020 season.
As a community we are yet to dive into the new solutions that have been released. You can find out more by inspecting and using the AWS SageMaker DeepRacer sample notebook. Just remember to clean up the resources (helper code available in the notebook) and stop the notebook when you stop to stop getting charged for it.
Thank you for this info, Tomasz. I am also noticing you can submit just one evaluation at a time, and only two training jobs in parallel. I checked for Time Trial. I did not have a chance to check other race types or across all 3 race types.