Researchers have developed a method for space robots to adapt to changing environments without needing external rewards. The approach uses a pre-trained model-based agent that learns to predict the reward structure within its latent space. This allows the robot to update its behavior based on imagined trajectories generated by an updated world model, even when it's experiencing hardware degradation. The method was tested across various simulated tasks, including planetary traversal and precision assembly. --- Why it matters: This matters because traditional reinforcement learning methods require a reward signal during deployment, which can be difficult or impossible in space due to the lack of external tracking systems. This approach provides a way for robots to adapt to changing environments without relying on external rewards. Source: https://arxiv.org/abs/2608.23452