Specialized vision-language-action (VLA) policies can acquire strong skills in specific domains, but merging those skills into a single shared policy has been a persistent challenge. A new paper on arXiv proposes a multi-teacher on-policy distillation method to address this, specifically for flow-based VLA models.
The approach builds on on-policy distillation (OPD), which transfers teacher knowledge to a student policy by training on inputs encountered during the student's own interactions. By extending this to multiple teachers, the method aims to consolidate diverse expertise without requiring the student to simply mimic any one specialist.
Because the abstract is brief, no experimental results are described. The contribution is framed as a technique for combining specialized policies, which could be relevant for developing more generalist robot and multimodal agents.