How anyone (yes, you too) can train their Robot
Well, we learned from the movie how to Train our Dragons :-).
But robots are a different breed. Unlike dragons, who are very smart, robots come as a blank book and are completely dumb.
Fortunately, smart people found a way how to train robots. The method is called Reinforcement Learning (RL) and is part of Artificial Intelligence (AI).
With RL we change behaviour of a subject by rewarding or punishing their actions. When a robot does something right, then it gets scores. When a robot does something wrong, well, then guy kinda gets killed.
At RL the system trains behaviour of the robot completely from scratch and repeats every successful step of the robot and eliminates every wrong one until robot learns to walk. Like a little baby who learnt to walk and trains moves trying to find optimal ones for initial success.
RL as a part of AI is complex process that need programming knowledge. At NextGrid we prepared environment where you can train a dumb robot to walk. For this you DON’T need any knowledge or experience, you just need to make a few clicks.
To make it even more insightful, we prepared two options when you can compare results regarding how much training cute robot gets. At first option robot gets a smaller amount of trainings (50,000 repetitions) and at the second 500,000 repetitions.
Not to spoil the surprise, try it by yourself by following simple step by step instructions below. Anyone can do it, it’s not hard. All you need is a Google Account.
Training cute robot to walk was a challenge at the 5th Edition of Deep Learning Labs in Warsaw.
We use model BipedalWalker (they gave cute robot a weird geeky name, I know) from OpenAI Gym.
So let’s train our cute robot now:
- Login into your www.google.com
2. Open training environment (click below)
3. Do a project copy for you
Go to ‘File’ and say ‘Save Copy in Drive’.
When your copy is created you see something like this.
Click ‘OPEN IN NEW TAB’.
Voila, now you are in your own project.
4. Run everything
Next step is simple, but takes few minutes to finish.
Just go to menu ‘Runtime’ and click on ‘Run all’.
After this system will start to set up and do training. Congratulations, you are part of AI now :-).
Now for a few minutes you’ll see a lot of happening to set up training environment. On right side of our browser you can scroll up and down to see different things.
Take a look at these icons. Icon like this means system is at this step.
Icon like this means that system is waiting to be executed.
Icon like this means things are finished so you are ready to go.
Don’t worry about all the technical mambo jumbo appearing. It’s highly interesting for experts but probably less so for non tech people.
5. Train your robot with 50,000 steps (simpler training)
Click on Train model 50k steps$ evaluate results (on left side of workbook)
After awhile on left side Click on Train model 50k steps & evaluate results
You guess it, here Is training going on for 50000 steps of training of our cute robot.
When it’s finished you can click the video and see the results.
6. Train robot with 500,000 steps
Now click on happening to train 500,000 steps on left side of workbook
It takes a while but when it’s finished you can see the results on video.
Yipiyaya, that’s it.
Now compare both video to see how training (number of steps) influences the result.
Advanced users can create their own section and add training steps and create a perfect walking robot.
Interesting fact:
Term Reinforcement Learning is taken from Psychology.
