Training in the Chamber


Oh wow, what an epic few weeks it’s been. Since last we spoke, we’ve been digging into the inner workings of AirSim, and working with our assigned sensors. My sensor, the distance sensor, was relatively straightforward to use, returning a value in meters of between 0 and 40, indicating the distance to the object in front of the car. I was able to build a simple script that took a measurement, accelerated about a foot, braked, and took another measurement. By doing so I was able to see the value decrease.

From there I went on to develop an interface for the distance sensor, enabling easy access to a boolean indicating whether a “distance threshold was breached” (a value below a minimum on the distance sensor was recorded). This was useful for my teammates, but where it really took off was as an interface to be used during training. I extended the car training model to treat any breach of the threshold as a collision, hoping to implement a “safety cushion” in the training model. While this might turn out to be a good idea, it might also have caused our simulator to get “stuck,” which caused me to have to restart training on my machine.

It’s not sensor integration time, but I’ve started training anyway on my end given that time is running out. After integration this weekend, at the next training epoch, I can replace the contents of dqn_car.py with our combined suite, and save data using the new suite to the existing model. I really hope this will work, but at worst we’ll lose a few days.

I’ve made some costly mistakes over the past few weeks, and I’m concerned about out ability to get this project in. We only have about 300k training episodes between here and the end of the term. This is worrying because the episode at which learning starts had originally been set to 300k, with the epoch ending at 500k. I have a feeling that changing these values was done at our peril, and may have unintended and unfortunate consequences. I’ll be reaching out to the Professor in hopes of finding a way forward.

Wish us luck! Sensor integration this weekend, followed by resumed training, if all goes well. Who knows, maybe the integrated car will train faster? A person can dream. Thanks for reading, and see you next week!

Print Friendly, PDF & Email

Leave a Reply

Your email address will not be published.