I think the bigger trap is to assume that a modular approach is the answer. For so many locomanip tasks the way we move our body depends so heavily on our environment + goals. I wouldn’t be satisfied to say “let RL handle task agnostic motion and just receive commands from a high level planner”.

For purely “move your body like this to location X” yeah I think current methods will scale. But once you start task conditioning (esp in human environments) I think it’s a lot less clear, both in terms of what behaviors we even want and how to specify them.

I’m using on shape as well since it’s free for me as a student!

Thanks for the insights, I’ll try!

once I have the basic idea and a little confidence I’m pretty good at self directing new projects, but starting is hard

honestly I think what would help me most is a list of increasingly difficult things to make that require more and more tools etc the same way intro coding classes give you assignments that use more and more features of the language.

meant to sleep 4 hours ago 🙃

Amazing post. One suggested addition, please please use data classes whenever possible instead of dictionaries. Raw dictionaries cannot be inspected without running the code and printing the keys. But data classes are basically self documenting! (And they can be typed).

I always feel burn out creep up at the end of the semester and the “take basic care of yourself” special continues to work wonders

But now I have to actually think about what I want my robot to do womp womp

Vs.

Now I get to think about what I want my robot to actually do!!

I know a “prove you’re not a robot” test hate to see me coming

Maia-core and flylist pilled

Or is that Sprigatito

Yoshi at the top of the food chain