What is the reason to train a Neural Network to estimate a task's success (i.e. robotic grasp planning) using a simulator that is based on analytic grasp quality metrics?

Isn't a perfectly trained NN going to essentially output the same probability of task success as the analytic grasp quality metrics that were used to train it? What benefits does this NN have with respect to just directly using said analytic grasp quality metrics to determine whether a certain grasp candidate is good or bad? Analytic metrics are by definition deterministic, so I fail to understand the reason for using them to train a NN that will ultimately output the same result.

This approach is used in high-caliber works like the Dex-Net2 from Berkeley Automation. I am rather new to the field and the only reason I can think of is computational efficiency in production?

  • $\begingroup$ Hi. Can you describe the difference between "analytic grasp quality metrics" and "analytic grasp metrics to determine whether a certain grasp candidate is good or bad"? Can you point us to a specific example of each of them? $\endgroup$ – nbro Jul 17 '20 at 11:59
  • $\begingroup$ Hi @nbro, I intended to refer to the same metrics in both cases. I've edited the question to make that clear. Analytic grasp quality metrics are deterministic and are based on finding, for example, hill-like shapes on the surface of the object that you want to grasp. An example of that is the paper I referenced in the question. The opposite of analytic success criteria would be empirical (data-driven) methods that determine success from pre-annotated human labels or physical trials with a robotic arm. An example would be this (Princeton-MIT). $\endgroup$ – EmVee Jul 18 '20 at 11:21

Your Answer

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

Browse other questions tagged or ask your own question.