Trustworthy and deployable foundation models for embodied and physical AI

Foundation models, including large language models (LLMs), vision-language models (VLMs), and vision-language-action (VLA) models, are rapidly entering the field of robotics and physical AI. By unifying perception, language understanding, planning, and action within single, scalable architectures, these systems present a significant opportunity for robotic agents to operate across diverse real-world tasks. Yet, unlike disembodied applications in natural language processing or image generation, deploying foundation models in physical systems places them under constraints that have no counterpart in digital environments: strict latency requirements, physical irreversibility of actions, sensor noise, mechanical variability, and the ever-present risk of harm to people or infrastructure. Meeting these constraints requires not only algorithmic innovation but robust systems engineering, a dimension that remains underexplored relative to the rapid pace of model development.This Research Topic addresses a critical challenge in the field. While foundation models have demonstrated remarkable capabilities in simulation and benchmark settings, translating these advances into reliable robotic systems for real world deployment remains an open problem. Current approaches often struggle with changing deployment conditions, complex physical interactions, limited computational resources, and collaborative scenarios involving humans and multiple agents.This collection brings together emerging solutions to three interconnected challenges. First, it explores how foundation model driven robots can operate safely and predictably in open environments. Second, it examines approaches for enabling deployment under real world computational and communication constraints. Third, it advances rigorous evaluation frameworks that assess trustworthiness, robustness, and reliability, moving beyond task performance alone to capture the requirements of embodied physical AI systems.Contributions focused on real-world robotic systems and deployment validation are especially encouraged. Contributions addressing the design, optimization, and integration of robotic systems and their components to enable trustworthy deployment of foundation models are also welcome. Areas of interest include, but are not limited to:• Safety-aware planning, constrained control, and failure detection in foundation-model-driven robots• Real-time inference, edge/cloud deployment, and hardware-aware model adaptation• Cross-embodiment and cross-domain generalization under deployment constraints• Sim-to-real transfer strategies and robustness validation protocols• Federated, privacy-preserving, and communication-efficient robot learning at scale• Benchmarking frameworks and trustworthiness metrics for embodied foundation models• Human-robot teaming under model uncertainty and transparency requirements• Real-world applications in service, industrial, assistive, and autonomous robotic systems• Hardware-software co-design and integration of sensing, actuation, and onboard computing for reliable foundation-model-driven robotic systems. Topic editors: Kongyang Chen, Jiaxing Shen, Mahdi Tavakoli Source: https://www.frontiersin.org/research-topics/83405/trustworthy-and-deployable-foundation-models-for-embodied-and-physical-ai