New AI Trick Makes Robots Faster and More Accurate

Robots that can hear a command, look around, and physically act are getting smarter—but they face a big speed-versus-accuracy tradeoff. If a robot replans its next move after every single action, it makes better choices but demands so much computing power that real-time use becomes impossible. If it plans ahead for many actions, it runs fast enough for the real world, but accuracy often drops.

A paper from researchers at Qatar University and collaborators proposes a clever fix. They replaced a heavy component of the popular SmolVLA robot brain with something called Mamba—a newer, lighter AI architecture. Imagine the difference between a mega-sized textbook and a well-organized cheat sheet: Mamba delivers similar knowledge but uses far less memory and speed.

The results are impressive. When the robot executed 50 actions before replanning—the scenario closest to real-time use—the Mamba-powered version succeeded 7.8% more often than the standard transformer-based approach. With 25 actions, it was still 3.7% better. And when the robot did stop after every action, the Mamba variant matched the old accuracy while shrinking the model's total size by 24%.

What does this mean for you? Cheaper, faster, and more reliable robots that can work in homes, hospitals, and factories—without needing a supercomputer behind them. The research is an early step, but it suggests that efficiency wins aren't just about cost; they can also make robots genuinely better at completing tasks in messy, real-world environments.