Robots that see and act need both a brain (the model) and a support system (the harness) that provides tools, memory, and error recovery. The most capable brains are large and expensive to run. This work asked whether a harness improved for a large brain could be passed to a small, fast one.
The team improved the harness of a large robot model by having AI suggest and test changes automatically. Then they used the large model's decisions, including its mistakes, to train a compact 4B model to replace it in the same harness.
The small 4B model performed almost as well as the large one while using one-fifth of the compute time. This shows that expensive robot brains are not always needed; their know-how can be packed into small models that run cheaply.