In offline reinforcement learning (RL), an agent learns from a fixed dataset of past experiences without trying things in the real world. When the dataset consists of camera images, predicting long sequences of future events is especially difficult because small errors in each prediction snowball over time.
MAC-v2 learns a compact internal model of the world without trying to reconstruct the original images. This avoids wasting effort on irrelevant visual details. It also predicts several future actions at once, reducing the number of steps the model must roll out.
Across many visual and robotic benchmarks, MAC-v2 matched or outperformed existing methods. This shows that effective long-horizon planning from pixels is possible, bringing offline RL closer to real-world applications like robotic manipulation.