Physical Intelligence : Not Most Efficient but Effective way to accomplish a task
Robots are being designed to perform complex tasks in the most efficient way. They are now beginning to defeat humans at some highly intelligent and complex tasks, like in AlphaGo game or Chess. Some robots can even play table tennis and other physical sports in an agile and intelligent manner too. but when it comes to understanding some simple tasks, they still lag a lot behind human intelligence.
Sometimes have really abstract instructions for some tasks like 'spread butter over bread', 'pick up a fruit from the pile' or 'clean dishes'. While performing such seemingly un-intelligent tasks as per today's standard, we might not be working with the most efficient manipulation algorithms, but we can perform these tasks with a high success rate. To design the best control scheme for these tasks, determining the best cost function even for Machine Learning becomes highly difficult. Therefore, there lies a need to come up with a methodology to design effective control systems for such tasks.
Professor Sangbae Kim of Biomimetic Robotics Lab at MIT, explains this during his presentation at GRASP on Robotics session at GRASP lab, UPenn.
If we take an example of subconsciously picking a fruit from a pile, we will not go on to optimize our motion, path, select the easiest apple to pick, and spend our computational resources on that. In my opinion, we will reach directly to the pile and while we touch many apples, we will pick one with which we will have the best confidence of picking it up successfully. We will determine this confidence through our estimation of the amount of grip we have on that object, probably through the contact force measuring, and we will pick the apple.
As humans, we have multilevel cost functions defined. We will probably take into account points like;
Time taken to accomplish the task (including behind the planning)
Efforts required to take the task
Potential Life risks
Other emotional factors
As the task of picking an apple doesn't take a lot of effort or energy. Thus, rather than doing overanalysis on the best way to pick the apple, I guess we prioritize the time spent on performing that task. So we will reach out to the pile right away, and will pick the apple ensuring the confidence of succeeding in the shortest time.
Comments