Towards Superior Cross-Domain Adaptability in Embodied AI: AMultimodal Adaptive Fusion Mechanism with Meta-Learning

Keywords: Embodied AI

Abstract:

This paper proposes a multimodal adaptive fusion mechanism for embodied AI, integrating physical
interaction experiences with human knowledge through environment-specific weight adjustment and meta
learning-based consistency validation. We hypothesize that this approach achieves superior cross-domain
adaptability and generalization compared to single-modality learning methods. Experiments on AI2-THOR
and RoboSuite benchmarks show significant improvements in success rates (up to 88.4% in target domains),
reduced generalization gaps (down to 4.4%), and faster adaptation (2-3x efficiency over baselines). These
f
indings validate the framework’s efficacy, offering insights for advancing embodied AI in robotics and
autonomous systems.

Submission Number: 261