nod. I think it was @jas0n who did one of the earlier requests…
He is trying to use physics… for optimizing the packing of various little mesh… into a fixed-sized container. Essentially, he is building a system that answers the question: Can a robotic manipulator arm… “learn” how to optimally pack a crate… with its “hand”… via learning from “bumping” that happens on the inertial sensor in its hand. It needs to “feel” a physics collision… in the robotic packing hand… to know when to release a container, or push a container, or turn a container, etc, etc. Phew. Tough.
Optimized packing of containers. FUN area of study and hobby. JasOn considered using physics collisions as sensors, with user or computer dragging… as his “mover”.
I know what everyone is thinking, here. Give me a list of ALL available sub-shapes sizes (items to pack)… and give me a container size… and “virtually” pre-pack the container… with a SERIOUS algorithm which analyses sub-container boundingBox dimensions. The robot would know how to pack the box… way before it started dragging items into the crate. (a dual layer thing - think it out, first, then do it.)
A physics-active “learn as you experiment” robotic arm “packer” would be pretty fun to watch. 
It might take DAYS to pack the container… MONTHS… it carefully watches onCollides and isSleeping on every PB… learning from its hand sensors… adjusting game-plans… unpacking and repacking - over and over. hehe. NUTS! I don’t think this is what JasOn has in mind… but… fun. I think we would find… pack with human doing best they can… ships in 10 minutes… WINS OVER… pack with robot AI-learning… ships in 10 months.