Pinned post
No prior data. No rewards. No pretrained models.
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 🧵👇
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 🧵👇
0:14