Which two instructions in the "programming language" of the 2011 HW would be the most difficult for robots to follow?
Well, though there were several candidates, finally, we came up with instruction 2 and 7 would be most difficult.
In step 2, the pencil needs to be held with precise force. Greater force may lead to slipping of fingers over the pencil surface. Lesser force may lead to falling of pencil in step 4. Also, so that torque can be applied in step 3, there should be some distance (few millimeters) between M and T along p-axis. That distance also needs to be quite precise.
In step 7, forces between M and T have to skillfully applied such that the contact point acts as axle and the object doesn't falls down. The friction at the axle is also decided by force applied. It also needs to make sure that force by I-effector is applied perpendicular to p-axis, other wise the pencil may slide.
The robot following the learning paradigm as in Kalakrishnan is clearly gaining some expertise. Which aspects of the execution may be called implicit or automatic, and which aspects may be more explicit? What could be the "chunks" in this structure?
There were debates, what do we actually mean by implicit and explicit learning for robots. Going by definition of tacit/implicit knowledge for humans, it is something that is difficult to transfer from one human to other via language. But in robots, every bit of knowledge can be transferred from one robot to other via robot's language of 1's and 0's. We finally, understood implicit knowledge as knowledge that is not provide explicitly to it (though procedure for acquiring that knowledge may be provided). After implicit learning, if one robot passes that knowledge to other robot, then than knowledge is not implicit for second robot.
In the Kalakrishnan's demonstration, the trajectory of hand, and configuration of fingers is explicit learning. However, during trails, it learns how much force and torque to apply. The process of learning the amount of force and torque to apply is implicit learning.
Chunks can understood as low-dimensional embedding of high-dimensional space. In the task, robot has to learn force and torque for each finger as a function of time. If we see task from higher level abstraction as sub-task done each second/step, the sub-task done in each second/step can be regarded as chunk.
Comment on whether human learning may also be following similar "reward" based processes? Consider the learning process for the fire-fighting expert who knows how to fight complex fires.
We support the fact that human learning follows similar policy of reward and punishment.
A firefighter has to analyse fire and learn, where to douse, how long, with how much pressure, etc. During the training phase of firefighter, he may try different strategies and face the consequences. His ultimate reward would be extinguishing the complete fire, but smaller rewards (like extinguishing of partial fire) also helps him learn the skill. The outcome of the strategies or steps applied by firefighter causes the firefighter to decide whether to continue or abandon that strategy.