Conjecture—Inductive biases based in action and perception

Establish whether inductive biases grounded in action and perception are necessary and sufficient to enable effective learning and generalization in embodied agents across diverse tasks and environments.

Background

The authors propose that appropriate inductive biases for embodied intelligence should be rooted in action and perception, reflecting the agent’s interaction with a physical environment.

They frame this as a conjecture motivating research directions on how such biases can support robust learning under physical constraints.

References

We conjecture that the necessary inductive biases for embodied learning are based in action and perception, and we address this conjecture in section \ref{sec:inductive-bias}.

— From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence  (2110.15245 - Roy et al., 2021) in Section 1 (Introduction: Guiding Questions)

Two questions remain: what information and prediction horizon are needed for robust decisions across changing objects, scenes, and embodiments; and whether geometry, memory, prediction, and multimodal action modeling justify their costs over simpler alternatives under matched data and computation.

— Physical AI: A Survey of Models, Learning, Systems, and Open Challenges  (2610.21171 - Gao et al., 15 Oct 2026) in Section 3.5.3, Architecture Comparisons