Sim TEMP

I'll cover these 3 questions:

  1. Why care about simulators?
  2. Why differentiable simulators?
  3. What are system-level considerations that are relevant to robotics that have yet to be fully realized?

Hopefully you walkaway with a better idea of which simulator you should use.

Why care about simulators?

  1. Rapid prototyping and development of:
    1. robot systems/morphologies
    2. algorithms for controlling robots
    3. (potentially co-designing)
  2. Generating large datasets that are required for modern machine learning techniques that have led to breakthroughs in areas like signal processing, machine translation, CV, and NLP.

There are different modes for using a simulator.

For certain goals, the fact that many simulators struggle to model real physical characteristics accurately has been ignored for at least half a decade, but likely longer. We need to start focusing on using simulators for the right reasons.

  • accurately model the dynamics of the robot
  • accurately model the dynamics of an environment (sometimes as a part of the total controllable system)
  • identifying parameters that describe the dynamics of the robot
    • masses
    • friction coefficients
    • etc.
  • See EMOs talk and table

Different simulators make different assumptions.

As we discussed last week, CLM problem solving methods are the most common:
(theses are acceptable when we have this specific assumption.)

Which simulator is good for each of these modes?

Summary Table of simulators

Why differentiable simultors?