The team at Alphabet have used a reinforced learning program to teach the DeepMind AI how to do parkour. Reinforced learning (RL) is a common tool for teaching and guiding behavior by using a reward system. Basically good or desirable behavior gets rewards and undesirable behavior gets nothing.
The aim of the project was to investigate if simple rewards systems would also work in complex environments. A virtual parkour course was designed with steps, ledges, hurdles, and drops. The AI was initiated to understand the faster it moved across the terrain the better the rewards. Bonus rewards were added for more complex programs.
The AI navigates the landscape learning to move forward as fast as possible without “terminating.” It uses a trial and error system to learn the methods as fast as possible. Watching the videos you can see the DeepMind is using creative thinking to move around obstacles with efficiency, even if the resulting movements look somewhat clumsy.
Research can be applied to IRL robots
The research was presented in a paper from the Google AI offshoot titled “Emergence of Locomotion Behaviours in Rich Environments.” The work being done here will form the basis of how to teach real life robots to learn to navigate tricky terrain like stairs.
While watching the stick figure navigate, it is important to remember that all these movements, the jumping, climbing, stretching and running were all self-taught. These are the movements that the AI developed to complete the task. While they often look like a drunk sketching model trying to catch the last metro home, in fact, you are watching AI history. All the DeepMind geniuses have given the agent a set of virtual sensors that allow it to know where it is and the incentive to move forward.
[Image Source: DeepMind]
Parkour or free running is a mode of moving through landscapes that relies on the fast flow of forward movement. The aim is to move from one place to another using the most efficient path and movements as possible. Derived from military training, Parkour has been likened to a non-combat martial art.
Massive leap for future AI training
This is a huge step forward in understanding the way RL can be used to teach complex movements. The AI successfully learns difficult and robust movements all through reinforced learning. Previously, reinforced learning had been thought to produce only fragile learned behaviors that would be dropped when exposed to unfamiliar conditions. The paper explains, “Reward engineering has led to a number of successful demonstrations of locomotion behaviour, however, these examples are known to be brittle: they can lead to unexpected results if the reward function is modified even slightly, and for more advanced behaviours the appropriate reward function is often non-obvious in the first place.”
The team rose to this challenge and have proved that in fact reinforcement learning can be used to achieve rich and effective behaviors. What is next for the Parkour AI and its real world applications is very exciting.