Breaking Rigid Priors: Toward Context-Driven Embodied AI
Stella Yu
Warren Hall room 205A and via Zoom (see note below to request the zoom link)
Short abstract: How can we move beyond hand-crafted assumptions in building intelligent systems? Rigid priors hold back embodied AI: forcing perception to handle only predefined transformations, treating locomotion and navigation as isolated silos, and reducing perception to a single, context-free snapshot. I’ll present recent work on data-driven, context-adaptive approaches that move beyond these constraints, pointing toward agents that see robustly, act cohesively, and reason flexibly.
Long abstract:
A central challenge for embodied AI is moving beyond rigid priors—fixed assumptions that oversimplify perception, action, or representation. In perception, these priors appear as hand-crafted equivariant architectures or augmentation-heavy training, constraining generalization. In action, locomotion and navigation are often studied in isolation, assuming complex behavior can be modularized into separate skills. In representation learning, the prevailing view treats an image as an objective measurement to be compressed into a single, context-free feature vector, ignoring that perception is subjective, context-dependent, and shaped by top-down as well as bottom-up processes.
I will present recent work that replaces these constraints with data-driven, context-adaptive methods. For perception, a test-time canonicalization framework leverages foundation models to map inputs into visually typical views, enabling robustness across rotations, lighting changes, and day-night variation without retraining. For action, a trail-hiking framework drives integrative skill development in humanoids by unifying long-horizon navigation with fine-grained locomotion. For representation, a contextualized feature learning approach supports open ad-hoc categorization, dynamically discovering novel concepts from sparse exemplars and unlabeled data. Together, these directions show how embodied AI can advance by replacing rigid priors—about invariance, modularity, or objectivity—with scalable, context-driven learning, opening the path toward agents that see robustly, act cohesively, and reason flexibly.
—
To request the Zoom link send an email to jteeters@berkeley.edu. Also indicate if you would like to be added to the Redwood Seminar mailing list.