Content
@
0 reply
0 recast
0 reaction
July
@july
There's a sort of transition / implicit conflict going on around robotics: it's between the "old guard" hard-coding robotics systems (control, planning, perception etc) where you want to limit your problem space, and hard code a FSM into it vs "new wave" of AI/ML folks that are going the end-to-end route. The end-to-end new wave is like LeRobot - just RL learn everything team, zero-shot learning take it sim2real lmao lfg Also historically the ML folks would do projects, and it wouldn't work, and we'd revert to more hardcoding things. Also perception didn't work that great previously, and now it does really well (generally speaking - the bottleneck is still great data) also sensors are getting cheaper (i feel like I keep saying this each year, but holy guacamole lidar is great, and a good chunk of why co's like Unitree succeeding) Also benjie talks about it in his blog a little bit https://generalrobots.substack.com/p/ml-for-robots-hybrid-learned-vs-end
5 replies
2 recasts
32 reactions
July
@july
Also it takes a long time for new methodologies etc to seep in generally to legacy projects (inertia of hardware changes etc / testing) so I don't expect everything to be end-to-end, but there are a lot of teams looking into it, but as always, it is the future, but it is overhyped. So the near term is somewhere between this new way, and the old way. The old way is underrated while the new way offers potential (and is not totally fluff) I'm excited about both
0 reply
1 recast
6 reactions
