New Training Trick Makes Robots Smarter With Way Less Data

Imagine a robot that can see your kitchen, understand “pick up the red cup,” and actually do it. That’s a vision-language-action model, or VLA. Normally, teaching these robots requires enormous amounts of data — millions of examples of robots doing tasks. But collecting that data is expensive and slow, because a robot has to physically move through the real world. This is the biggest bottleneck in robotics.

The research team behind VLAct took a different approach. Instead of just feeding the AI more data, they focused on how it stores and connects what it learns. First, they trained the model on a wide variety of robot data — different robot bodies, different environments — then fine-tuned it for specific tasks. This creates a strong foundation that lets skills transfer from one robot to another, even ones it has never seen.

The results are striking. VLAct beat two major industrial robot systems, ABot-M0 and LingBot-VLA, with success rates around 83% and 93% on standard tests. Even more impressive, when tested on a humanoid robot it had never encountered, it used only 20% of the training data to outperform a baseline that had the full dataset. And all of this was done using just 16 GPUs — modest compute by modern AI standards.

Why does this matter? The paper shows that clever training can matter as much as raw data. If robots can learn more with less, they become cheaper to build and easier to update. That brings home assistants, warehouse automation, and even care robots closer to reality. The code and models are fully open-source, so other researchers can build on this work right away.