Robots often pair a large vision-language model (VLM) with a smaller action expert. Training the expert can update the VLM, potentially degrading prior capabilities. Some block the expert's influence (insulation), but the best approach is unclear.
The researchers compared blocking (insulation) versus allowing (coupling) the expert's updates to the VLM, and whether the VLM also learned action tokens representing target actions. They tested open-loop prediction and closed-loop task completion in simulated driving and manipulation.
Insulation was competitive only when action tokens were used. Without them, coupling improved prediction and task success. In closed-loop tests, coupling often helped even with action tokens. Thus insulation should not be the default; the choice depends on supervision and evaluation setting.