Effector Complexity Enhances Transfer Learning

ICLR 2026 Conference Submission15748 Authors

Keywords: Embodiment, Transfer Learning

TL;DR: A study on how embodiment shapes transfer learning: simple effectors train faster from scratch, but complex effectors yield greater improvements after transfer.

Abstract: Both biological and artificial embodied systems rely on effectors to interact with the world. How does this embodiment impact the way they learn? The role of embodiment in shaping learning dynamics is not well understood from a neuroscience or a machine learning perspective. In this study, we treat embodiment as a variable in artificial agents and study how changes in effector complexity reshape the dynamics of learning. Our hypothesis is that more complex effectors provide constraints that yield better transfer learning on new tasks, despite simultaneously posing a more complex control problem. We evaluated this hypothesis on area under the performance curve, and use time to sustained performance plateau as a parameter for task difficulty. Our results show that while a simpler effector excels when trained from scratch, a more complex effector yields superior performance after pre-training on another task. We further demonstrate that the improvement gained from transfer learning is greater for the complex effector. Our findings suggest that embodiment plays an important role in enabling efficient transfer, offering insights into the differences in learning dynamics between disembodied artificial systems and their biological counterparts.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 15748