Robots are often trained with a specific software controller that translates commands into motion. Different controllers make the same command produce very different movements. Training a robot on data from many controllers usually fails because a command that works for one is wrong for another.
This work lets a robot adapt. Before a task, the robot watches a few example commands to see how it actually moves under its current controller. A neural network uses these examples to understand the robot's behavior, guiding the robot's own actions for the task.
In real tests, this method improved task success by 33 percentage points over standard data mixing. It worked on controllers the robot had never encountered and on tasks requiring physical contact. This helps build robots that can handle diverse hardware setups seamlessly.