|Amazon's DeepRacer - Careful What You Wish For|
|Written by Mike James|
|Saturday, 01 December 2018|
This sounds wonderful. If only it was going to be available to order in time for the holiday season. However, it might cost you a lot more than the $400 price tag suggests. Reinforcement racing anyone?
Reinforcement Learning (RL) is big in AI at the moment. You don't tell the system what is right or wrong, you simply provide feedback in the form of a reward appropriate to the action and its result. The theory of RL isn't particularly difficult, but applying it can be. RL systems tend to be very slow to learn and training cycles go on and on and on.
Amazon wants to encourage us to get involved in AI, and RL in particular, and at this year's Re:Invent conference one big announcement was DeepRacer. This is described as a 1/18th scale race car - although what it is a scale model of isn't clear as it looks very strange:
That is the front view of the car and you can see the camera that is used to navigate round the track. You get the impression that Amazon's publicity department thinks the front is the sharp end from a lot of the publicity shots - see the photo at the end.
The hardware specs are fairly modest:
Notice that it uses a fairly small processor and not a huge amount of memory. The reason is that a trained RL models are comparatively light to work with. The software is used is quite interesting. The operating system is the well known ROS - Robot Operating System running under Ubuntu. The vision system is Intel's OpenVINO which is OpenCV/TensorFlow based and can make use of hardware acceleration provided by GPUs or by Intel's Neural Compute Stick. However there does't appear to be any hardware acceleration in DeepRacer - perhaps using some of the USB ports for a few Neural Compute Sticks might give your DeepRacer the edge over the rest.
The hardware is going to cost around $400 when it is available, sometime next year, but if you order it now you can get it for just $249 - a bargain. And yes, many will want to buy one and join in the DeepRacer league:
All great fun but...
Reinforcment learning isn't easy and, despite the reassurance that you only need basic Python programming, my guess is that you will need rather more. There seems to be some custom software for the device. A racetrack simulator is provided to train and test the model. The four-step "Getting Started" states that you can create a model, train on a simulator, submit to a leaderboard to show how well you are doing and then deploy the model to the model racing car to try it out for real. In this case the devil is in the detail - train on a simulator.
When you look into it a little more DeepRacer is based on Amazon SageMaker for the deep learning models, the new AWS RoboMaker for the simulator, Kinesis for the video streams and also S3 storage plus CloudWatch for the logs. So you are going to be heavily committed to AWS usage and while these services have free tiers you are likely to exceed these and end up spending real money. SageMaker, for example, has a 2-month limited free period and after this you pay according to the size of the machine instance you are using. A system that is described as a moderate performer will cost $0.25 to $0.50 per hour and much more if you want to use GPUs.
It could be that Amazon is planning to create a special deal on compute time, but given the amount of computer hardware and time you have to throw at RL to get it to work - DeepMind took 5000,000 minutes of compute time using GPUs to learn to play an Atari games - I don't think it can be that generous.
So DeepRacer looks fun, but it could well cost you much more than the $400 headline price.
or email your comment to: firstname.lastname@example.org
|Last Updated ( Saturday, 01 December 2018 )|