Researchers found that a small AI model for robot manipulation performed better when trained on scrambled geometry data than on the real thing.
The study tested a 0.8-billion-parameter hybrid language model adapted for robotic manipulation, tuning it with just 6.2 million trainable parameters. Researchers trained six versions on three LIBERO-Spatial manipulation tasks and evaluated them over 540 held-out rollouts across three seeds. Conditioning the model on real object and goal geometry produced 28.9% success, worse than the 36.7% achieved when those same geometric signals were shuffled at random during training, and only slightly better than the 24.4% baseline with no geometry at all. A separate approach using timing signals instead of spatial coordinates scored just 11.1%, with one run failing to converge entirely.
The result cuts against a common assumption in embodied AI: that giving a model accurate, structured information about its environment should make it more reliable, not less. It also points to a deeper generalization problem: a robustness test in the same paper found a policy using relative coordinates held onto 7 of 10 successes when its reference frame changed, while every vision-based policy tested collapsed to 3 successes or fewer out of 20 after the target object moved just 5 centimeters.
In an industry racing to bolt language models onto robot arms, a paper showing the 'obvious' engineering choice doesn't even beat noise is a useful reality check.