In a talk at Stanford today, a graduate student presented his work on autonomous cars, robots, and helicopters. It’s amazing to see the four-leg robot walk over uneven surface without falling, and the unmanned helicopter doing in-air flipping and circling acts.
While the topic of the talk was about reinforcement learning, the real issue is dynamic decisions for a sequence of actions. A helicopter flies through space, but also over time. A robot has sequence of actions to take. It needs to react optimally to the spatial environment, but also take into account of sequential results of each step. Through learning (training with space features), and optimization over timed sequence, these robots (including vehicle and helicopter) perform their task very well.
How much is the success due to reinforcement learning, and how much due to the actual tedious engineering on the robots? I guess the later part accounts for 80%. Of course, there exists a way to scientifically measure the ration of contribution. For the practical purpose, people are much more interested in the walking and flying robots, than the abstract algorithm.
Engineeing is the mundane work to make something happen. I used to stay away from engineering, but now I embrace it. Only by being hands-on, I truly understand the real problems. Of course there exists a balance of utilizing other forces beyond one’s own. Then one has to build good alliance. Regardless how we do it, the goal is to make things happen, and happen fast.
