Robots with mismatched sensors can now adapt on the job without retraining

When robots with different sensors run the same pretrained software, performance drops sharply — a deployed team with one camera and one lidar fails where a team of identical robots succeeds. This paper shows a method that lets each robot learn to translate its own sensor data into the format the shared software expects, using only the team's final success or failure as a signal, no fine-tuning required.

Deployed robot teams are always mismatched — different platforms, different sensors, different failure modes. Until now, that meant either retraining the whole system for each new hardware configuration, or accepting degraded performance. This method keeps the core software frozen and lets individual robots adapt their sensor interpretation at deployment time, which means teams can stay heterogeneous without paying a performance cost. The catch is it only works if you can measure whether the team succeeded or failed as a whole, which works for warehouse logistics and search-and-rescue but not for tasks where you need real-time feedback on individual robot performance.