Very excited to share @jonghae717's new paper on goal-conditioned RL!
I've enjoyed working on this over the last 9 months because there are many mental models that can be used for understanding the same (simple) approach 🧵
I've enjoyed working on this over the last 9 months because there are many mental models that can be used for understanding the same (simple) approach 🧵
No prior data. No rewards. No pretrained models.View quoted post on X
Can an RL agent learn to solve extremely long-horizon tasks (e.g., 5-cube stacking) from scratch?
We show it can! C-Slerp learns hyperspherical contrastive representations and infers waypoints from them in closed form 🧵👇