Study on Proprioceptive Encodings for Robotic Manipulation

A new study from arXiv (2605.13067) investigates strategies for encoding a robot's proprioceptive state to improve manipulation robustness. The research addresses the gap between training and inference conditions in end-to-end robotic policies, particularly for robots with moving frames of reference. Through systematic evaluation of joint representations, the authors find that a simple episode-wise relative frame offers the best trade-off between task performance and robustness, outperforming baselines in extensive real-robot experiments in a realistic test environment. The work aims to improve zero-shot generalization to unseen test conditions.