Ohh exciting, would be keen to help out, anything in particular you’re thinking of implementing?

I’m surprised Gemini doesn’t push post-training super hard too given deepminds presumably very strong RL infrastructure.

Have you seen any approaches to rlhf that include hierachical/curriculum learnin? It would be interesting to see the trade offs with this 🤔

Full blog at: aos55.github.io/deltaq/
#robotics #machinelearning #reinforcementlearning

I've also open-sourced some code - SAC and IL implementations for Fetch Pick-and-Place, with pre-trained models on HF. If you're interested in contributing, the repo is open for PRs!

github.com/AOS55/RLFoun...

Currently covering:

- Why "simple" tasks are often hard for robots
- Learning from demonstrations
- Reinforcement learning in robotics
- Common supervised learning problems in robotics

Coming soon:

- Sim2real transfer
- Modern approaches
- Real-world application

∇Q, aims to break down why robotic learning is different from standard ML. Sure, we can train vision models on millions of images, but why can't we do the same with robots?