Robots Learn Better When Their Brain Cells Keep More Detail

Imagine trying to learn a new skill, but every time you practice, your brain rounds off the details. That's what's happening with many AI models that control robots. These models, called vision-language-action (VLA) models, use a number format called bfloat16 (bf16) that saves memory but loses precision. When fine-tuning these models, developers often skip a step that keeps a high-precision copy of the weights. As a result, most of the small updates made during training get erased—up to 79% of them. This means the robot barely improves, even after lots of training.

A new study tested this on two popular robot models: SmolVLA and π0.5. They found that using a full-precision (fp32) master copy during training dramatically improves performance. For example, on a task where robots pick up objects, success rates jumped from 4.3% to 49.3% for SmolVLA and from 5.1% to 80.9% for π0.5. Even at default settings, the improvement was significant. The researchers also checked 12 open-source robot codebases and found that 9 of them use this flawed training method by default.

The fix is surprisingly simple: keep the trainable weights in fp32 during training. For SmolVLA in the LeRobot library, it's a one-line change. This doesn't require more data or complex algorithms—just a tweak to how numbers are stored. The study shows that this small change can make robot learning much more effective, saving time and money. It also highlights a hidden problem in many AI projects: shortcuts in precision can silently undermine progress.

For anyone using or developing robot AI, this is a wake-up call. Checking your training setup could be the difference between a robot that barely works and one that performs reliably. As AI becomes more integrated into physical tasks, from warehouse robots to home assistants, these details matter. The good news is that the solution is easy and doesn't require a PhD to implement.