Chris Paxton#737
Supervised learning is kinda lame. I've always found RL the cool part of ML. It's what can make machines feel truly alive!
So far RL has had little practical use in robotics, but not anymore! The next version of openpilot's driving model will be fully RL trained.
Users are excited about Openpilot switching its driving model to full reinforcement learning training because they find the change super cool and are eager to see further developments.
@___Harald___ ngl RL making it to a production driving model changes the game a bit
what makes u pick RL over imitation for this specific case?
@luminxbt We have a lot of issues with controls, since openpilot works in so many different car models with laggy or inaccurate steering APIs.
So we frequently get issues of bias, and oscillation. With RL you can train in closed loop and penalize those things. Much harder with IL.
@___Harald___ > So far RL has had little practical use in robotic ?????????
@___Harald___ @comma_ai ooh I 100% agree and am excited to see this! are you going to be using any form of inverse rl from collected data? Or is this more sim2real style?
@___Harald___ Hey Harald, how is the lateral performance on RL looking like for the Bronco you all have in your fleet? It’ll be nice having good Ford performance out of the box in OpenPilot without the need of a fork.
@___Harald___ @comma_ai super cool! Is there a blog post?
@___Harald___ Hype is high, execution matters most always
@___Harald___ Openpilot RL sounds interesting, let's see results