If you were at the last RLC NeurIPS event, you know it's not to be missed. Invite any RL researcher you know, no invite-only parties here.

Robotics manipulation to be specific :)

Can someone point me to any paper that uses RL on real-life (image-based) environments without sim2real/imitation learning? For good reasons I am told that this is pretty common but I’ve only found a handful of papers (CQN, QT-Opt, SAC-X)

Hi Csaba :)

Streaming Deep Reinforcement Learning Finally Works Natural intelligence processes experience as a continuous stream, sensing, acting, and learning moment-by-moment in real time. Streaming learning, the modus operandi of classic reinforcement learning ...

Streaming Deep Reinforcement Learning Finally Works, by
M. Elsayed, G. Vasan, A. R. Mahmood, is one of those papers I wish I had written 😅

This paper seems to allow us to do RL with NNs as it should have always been done. Everyone should read it!

arxiv.org/abs/2410.14606

I also think many robot learning papers are overclaiming what they can do… the paper task setup can be very easy in comparison to real production systems even for pick-n-place… but it’s hard to see this difference in the presented videos (if any)

I feel like many works do have experiments on sim but they don’t seem to transfer to real life (not in terms of sim2real but applying the same algo). I wonder how much of it comes from delays or us being overprotective of the robot in real life. Maybe evals need to include these components.

I thought that was just me 😅 was trying it on an uncluttered single item picking task

Perhaps it’s “necessary” to have it as a baseline (arguably TD3 is just as important imo but SAC seems to be more commonly used), and it’s hard to convince people which one is stronger than SAC. I think recently there are few e.g. ACE at Neurips. Generally feels like a popularity game to me

Let me try, we’ve been very quiet historically 😂

Can I get added please :)

Please see me :)

Please add me, I’ve been doing robot learning with RL/IL!