Model-based vs. Learning-Based, which is better?

I had the opportunity to attend a seminar on robotics hosted by three distinguished MIT PhD students today. It is about motor coordination, manipulation, and physical interaction. There are two topics that lift my spirits. The first one, the conventional control theory, is quite elegant and unveils the beauty of mathematics and physics. The second thing is that two different control factions are emerging. Then, I will talk shortly about these two topics.

conventional control theory

Federico Tessari introduced the humanoid robot first and compared the human body with the Boston Dynamics robot. Apparently, robots have more “significant” performance, which means fast motion speed, fast electrical-mechanical speed, etc. But why is the robot's real performance so bad compared to humans? So here comes the paper “On Human Motor Coordination: The Synergy Expansion Hypothesis” . I have nothing to comment about this hypothesis. I only want to say that I found the numerical example quite graceful compared to the RL, in which we don’t know exactly what’s happening in the neuron networks. I can still remember the experience when I was refining the continuous control result using DDPG. The only thing I can do is adjust the reward function, the gradient of policy, or adjust the model structure based only on my intuition and past experience. So the RL really be a good solution for control? Then comes the next topic, model-based control vs. learning-based control.

Model-based vs. Learning Based

An instructive discovery is that I am not the only one struggling with which direction is better. Even the postdoctor at MIT doesn’t know which one is better. I can roughly recall the original sentence: “In the US, there are also two factions that are diametrically opposed to each other. ” The model-based control, such as IK and MPC, is stable for the control results, but when the model becomes sophisticated, the possible solution is hard to find. For the learning-based control, such as the RL, the problem mainly lies in the black box model. As a product, the customer can not endure the result is not under control. I believe such a dispute may continue for a long. And no one knows the final results. The only thing I can do is choose the direction for which I have the most passion. May the passion lead me to the next step.