ForeAct: Steering Your VLA with Efficient Visual Foresight Planning

Researchers introduced ForeAct, a visual foresight planner that dramatically improves robot performance. It generates a high-quality 640x480 image of a predicted future observation in just 0.33 seconds on an H100 GPU. This allows Vision-Language-Action models to focus on motor control instead of high-level reasoning. In tests across 11 diverse real-world tasks, it achieved an 87.4% success rate, a massive 40.9% absolute improvement over the baseline.