0
$\begingroup$

I read a paper from 2015, "Structural bootstrapping - A novel, generative mechanism for faster and more efficient acquisition of action-knowledge ", which introduces a concept called, "Structural bootstrapping with semantic event chains and dynamic movement primitives," which confused me a little bit.

According to my knowledge a robotarm is controlled by a PDDL-like planner. The PDDL file is a "qualitative physics engine" which can predict future events. The paper says the "qualitative physics engine" consists of dynamic movement primitive (DMP) which are learned from motion capture data.

My question is: How can a DMP be used for simulating physics?

$\endgroup$
1
$\begingroup$

I don't know that the DMP method actually simulates physics as much as it copies what it saw, which was an actual, physical system. Note that "qualitative" means essentially "looks like", where "quantitative" means "numerically".

Qualitative physics simulations are those that look like real physics, but they are not. If I hold a dart in my hand and walk it across the room as though it were in the air, it would look roughly correct, but of course it would not be numerically accurate. So again, I think they're copying what they see and treating it as "good enough."

| improve this answer | |
$\endgroup$
0
$\begingroup$

The DMP method copies a trajectory instead of simulating physics. The guys in the paper store the recorded trajectory into a database as a matrix and reproduce it without modification on the robot. The event-chain matrix is on page 5 at Learning the semantics of object-action relations by observation But, in the paper Execution of a Dual-Object (Pushing) Action with Semantic Event Chains is on page 4 a learning model used which is able to determine future result of an action (chapter "Defining Spatial Anchor Points"). My impression is, that a little bit of planning is in there. The author named it "model-sec" (Model Semantic Event Chain).

| improve this answer | |
$\endgroup$
  • $\begingroup$ "The guys in the paper store the recorded trajectory into a database as a matrix and reproduce it without modification on the robot." I think they do attempt to modify it, based on evaluating what parameters are variable. See page 22 and the explanation of the stir/wipe sequence; I think it talks about the starting frame or key point being fixed by the end of the previous motion, leaving "only the adjustable parameters" or something similar available to the user to set. $\endgroup$ – Chuck Aug 13 '16 at 17:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.