Can Model Arithmetic merge policies from distinct tasks?

Determine whether the Model Arithmetic weight-space merging strategy can successfully integrate policies that were independently trained on distinct robotic manipulation tasks (rather than only on complementary subsets of a single task’s demonstrations) in order to advance general-purpose robotics.

Background

The paper proposes Model Arithmetic (MA), a weight-space merging approach that combines checkpoints trained on complementary subsets of demonstration data to mitigate coverage deficiencies in imitation learning. Empirically, merging subset-trained policies improves success rates and throughput compared to training on the full aggregated dataset.

While MA is shown to be effective when merging models trained on subsets of the same task, the authors explicitly note that it is unknown whether MA can integrate policies trained on entirely different tasks. Establishing this would be a critical step toward building a single, general-purpose manipulation policy via weight-space synthesis rather than data or architecture scaling.