Overlap between expert and self-supervised tasks

Determine whether adding self-demonstrations on expert tasks contained in pretraining improves robustness or sample efficiency, and characterize which degrees of task overlap between expert-supervised and self-supervised data are beneficial or harmful.

Background

In the experiments, self-supervised tasks are generally distinct from the newly expert-demonstrated tasks, allowing the method to preserve pretrained behaviors while learning new skills. The paper does not resolve whether replaying self-demonstrations for tasks that also receive expert supervision would provide additional robustness or sample-efficiency gains, nor which amount or type of overlap between the two data sources would improve or degrade performance.

References

Does adding self-demos on expert tasks (if pretraining contains the same tasks) improve their robustness or sample efficiency, and what kinds of task overlap between ES and SS is helpful or degrades performance?

— Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation  (2608.19490 - Garg et al., 19 Aug 2026) in Section 6, “Future Work and Limitations”