3 ms·

sim2real research for robotics is super exciting. For all the work Boston Dynamics has done making carefully hand-tuned locomotion controllers, recent work has

• jcims 8y ago
  That’s super interesting, thank you for sharing. Lex Fridman has a recording of the CEO of Boston Dynamics on his podcast. It was from a lecture Q&A at MIT I believe. And in there he talks about the control challenges as though they are likely the bigger problem to tackle vs. the mechanical (which seems obvious on one hand but interesting to hear him say). In an off-handed comment he said how hooking the mechanics up to a person (so they could run the robot in a fly by wire way) shows just how fast the machine is and how much the software/sensor/controller stacks are slowing them down. He also discusses their simulation capabilities a bit. It definitely sounds like there is some software training going on, but this obviously cranks the fidelity up quite a bit. Would be cool to see how they train it.
• tnecniv 8y ago
  This is neat. Any idea if in the paper they address how well the network works if you swap out the robot for another one of the same model / if the parameters of the physical robot change?