Fengze Jia

Mathematics undergraduate at The Ohio State University, graduating May 2027. My current research asks why a compressed robot policy that passes offline evaluation can still fail in closed loop, and what repairs it.

Applying to MS programs in robotics and machine learning for Fall 2027.

Anatomy of a Closed-Loop Collapse: A Causal Case Study of a Compressed VLA Policy

Poster at the IROS 2026 Workshop on Building Scalable Infrastructure for Robot Learning (ScaleInfra), Pittsburgh, October 1, 2026.

I distilled Octo-Base, a 12-layer generalist robot policy, into an 8-layer student. The student passes the offline validation checks its model family provides, then succeeds in 0 of 72 closed-loop trials of a simulated WidowX pick-and-place task, where the teacher succeeds in 40.

Episodes reaching each stage, teacher versus 8-layer student Out of 24 episodes. Moved the eggplant: teacher 20, student 18. Grasped it: teacher 20, student 11. Held the grasp: teacher 19, student 11. Placed it in the basket: teacher 15, student 0. Teacher, 12 layers Student, 8 layers Moved 20 18 Grasped 20 11 Held 19 11 In basket 15 0
Episodes, out of 24, that reach each stage of the task. The student falls behind stage by stage, then never gets the eggplant into the basket.

Four standard fixes fail under matched controls. Replacing half of each training batch with successful teacher rollouts from the deployment environment restores parity with the teacher on held-out episodes: 18 of 36 against 17 of 36.

Cite this paper
@misc{jia2026anatomy,
  title         = {Anatomy of a Closed-Loop Collapse: A Causal Case
                   Study of a Compressed VLA Policy},
  author        = {Jia, Fengze},
  year          = {2026},
  eprint        = {2609.23048},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.23048}
}

Other research

Talks and posters

Service

Reviewer, IROS 2026 ScaleInfra Workshop.

Contact