A new reinforcement learning algorithm, called Contrastive Planning, demonstrates that complex manipulation skills can emerge without any external supervision. The agent learns to build towers of up to eight cubes from a blank slate, with no prior data, reward signals, demonstrations, or pretrained components.
The method relies on self-supervised hyperspherical contrastive representations, which appear to give the agent a structured way to plan and explore. By organizing experiences in this representation space, the algorithm can discover useful behaviors purely through interaction.
Because the source is a single arXiv abstract, the full details of the architecture and training procedure are not yet available. The result is notable for showing that tower-building—a task often used as a benchmark for physical reasoning—can be solved without any task-specific supervision.