Deploying AI models into the physical world requires robots to react almost instantaneously to environmental shifts. While current Vision-Language-Action (VLA) models—which translate visual inputs and text instructions into physical movements—can generate smooth trajectories, they often suffer from significant reaction lag. This delay occurs because standard systems wait to calculate an entire sequence of movements before starting the first one, a bottleneck that makes high-speed tasks like sports nearly impossible for AI.

Researchers from The University of Hong Kong have introduced FASTER (Fast Action Sampling for ImmediaTE Reaction) to bridge this gap. By rethinking how models group sequences of actions together (action chunking), the team developed a Horizon-Aware Schedule. Instead of treating every step in a planned movement sequence with equal importance, this method prioritizes the very first action. It compresses the initial calculation—the process of refining data from noise (denoising)—into a single step, allowing the robot to begin moving ten times faster than previous leading models.

Crucially, this speed boost does not sacrifice the quality of the robot's overall path. The system maintains a streaming connection between the AI's processing unit and the robot's physical hardware, ensuring smooth execution even on consumer-grade hardware. In real-world demonstrations, the FASTER framework enabled a robot to handle the chaotic dynamics of table tennis, proving that general-purpose AI policies can finally match the split-second demands of the physical world.

Deploying AI models into the physical world requires robots to react almost instantaneously to environmental shifts. While current Vision-Language-Action (VLA) models—which translate visual inputs and text instructions into physical movements—can generate smooth trajectories, they often suffer from significant reaction lag. This delay occurs because standard systems wait to calculate an entire sequence of movements before starting the first one, a bottleneck that makes high-speed tasks like sports nearly impossible for AI.

Researchers from The University of Hong Kong have introduced FASTER (Fast Action Sampling for ImmediaTE Reaction) to bridge this gap. By rethinking how models group sequences of actions together (action chunking), the team developed a Horizon-Aware Schedule. Instead of treating every step in a planned movement sequence with equal importance, this method prioritizes the very first action. It compresses the initial calculation—the process of refining data from noise (denoising)—into a single step, allowing the robot to begin moving ten times faster than previous leading models.

Crucially, this speed boost does not sacrifice the quality of the robot's overall path. The system maintains a streaming connection between the AI's processing unit and the robot's physical hardware, ensuring smooth execution even on consumer-grade hardware. In real-world demonstrations, the FASTER framework enabled a robot to handle the chaotic dynamics of table tennis, proving that general-purpose AI policies can finally match the split-second demands of the physical world.