ForeTime-VLA Improves Robot Grasping of Moving Objects

Manipulating objects in motion, such as those on a conveyor belt, presents a significant challenge for robotic policies, as it requires anticipating future contact events. Traditional vision-language-action (VLA) policies often rely solely on current observations, limiting their ability to react effectively to dynamic environments. While world action models (WAMs) can predict future dynamics, deploying them at scale or explicitly imagining future frames is computationally expensive. ForeTime-VLA addresses this by introducing a dense policy that distills a future-aware, action-equivalent representation from a pre-trained, frozen Fast-WAM-derived teacher. This "causal future-token distillation" allows the policy to learn predictive dynamics offline while remaining efficient during inference. The system uses an eight-frame history encoder to predict future video latents, manipulation phase, and time-to-transition, which then condition the VLM prefix and action expert. In real-robot evaluations on a conveyor-belt dataset, ForeTime-VLA achieved significantly higher grasp success rates—81.1% for stationary and 58.9% for slow-moving objects—outperforming the next-best reference by substantial margins, especially at faster belt speeds. This demonstrates that distilling future-aware information causally is an effective strategy for enhancing dynamic manipulation capabilities without the overhead of deploying a full world model.