Limits of imitation learning for high-precision insertion

BSc thesis, ETF Robotics Laboratory, University of Belgrade (2026). Advisor: Prof. Kosta Jovanović. Code · Policies on Hugging Face · Eval results

How far can imitation learning go on a task that needs millimetre precision? I studied this on fiber-optic (SFP) connector insertion with a UR5e arm in NVIDIA Isaac Lab.

• Built the task environment (AIC-Port-Insertion-v0) and a scripted oracle that generated 30,290 successful expert demonstrations.
• Trained ACT and Diffusion Policy (LeRobot) with an auxiliary phase-prediction head, at 1k, 10k and 30k demonstrations.
• Evaluated closed-loop over 150 episodes per policy, with automatic failure classification.

Findings. The main obstacle is lateral error, the sideways misalignment with the port. More data helps a lot: Diffusion Policy’s median lateral error dropped 3.5× (41.2 → 11.8 mm), but that is still far from the millimetre precision the insertion requires.