Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Robustness from failed generative rollouts

Determine whether fine-tuning on generative rollouts from failed episodes improves policy robustness.

Background

Many of the self-demonstrations used in the paper are unsuccessful or contain undesirable behavior, yet training on them still improves performance in several settings. This observation motivates an unresolved question about whether failed rollouts should be deliberately incorporated as training data to improve robustness, rather than merely tolerated or filtered.

References

Does fine-tuning on generative rollouts of failed episodes add robustness?

— 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”

Whether learning from both signals raises recovery beyond supervised repair remains open.

— World State Generator  (2609.24744 - Jeong et al., 21 Sep 2026) in Section 7, Discussion

BC is trained on demonstration futures; direct BC training on generated futures remains untested.

— Reliability-Aware Future Conditioning for Temporally Robust Robot Manipulation  (2610.11956 - Khoshnazar et al., 8 Oct 2026) in Section Limitations