The Physical AI Research (PAIR) Center pioneers the next generation of intelligent systems by bridging simulation and reality, advancing multimodal perception and reasoning, and enabling secure, trustworthy deployment in the real world. Through deep interdisciplinary collaboration, PAIR tackles the most critical challenges at the intersection of AI, physical environments, and human impact.
Research Focus Areas
Bridging Digital Worlds to Real World Generalization
Creating physics-consistent digital twins and generative environments to accelerate the development, validation, and safe deployment of AI in the physical world.
Goals: Overcoming Data Limitations and Closing the Sim-to-Real Gap
- Advance data-efficient learning techniques such as weakly-, semi-, and self-supervised learning, as well as active- and meta-learning, to enable robust performance on small, noisy, or imbalanced datasets.
- Develop physics-informed learning strategies that integrate real-world constraints and physical laws into the training/simulation process, improving generalization in dynamic environments.
- Leverage simulation and generative modeling to create realistic synthetic data, augment edge cases, and stress-test models under diverse and risky scenarios.
- Enable sim-to-real transfer through domain adaptation, novel software abstractions, domain-specific languages, and reinforcement learning frameworks with real-world fine-tuning and continual adaptation.
Perception, Reasoning, and Action in the Real World
Fusing multimodal sensing, semantic understanding, and control to enable AI agents that can reason and act with foresight in dynamic and uncertain environments.
Goals: Unified Understanding of the Physical World
- Build large-scale world models that integrate visual perception, spatial reasoning, and multimodal fusion.
- Develop temporal prediction pipelines that shift from reactive decision-making to foresight-driven planning and actuation.
- Bridge symbolic reasoning with continuous sensorimotor control for embodied intelligence in robotics, vehicles, and interactive systems.
- Co-optimize perception and policy under real-time constraints, with feedback from the physical world.
- Enhance our unified understanding of the physical world through integrated HPC and ML computational pipelines.
Safe, Robust, and Trustworthy Deployment of Physical AI
Ensuring physical AI systems are verifiable, trustworthy, and resilient to uncertainty, enabling safe interaction with people and environments.
Goals: Building Reliable Physical AI
- Embed formal verification, uncertainty quantification, and failure prediction into the AI development lifecycle.
- Design real-time guardrails and monitoring systems to detect and correct unsafe behaviors post-deployment.
- Promote explainability and interpretability, especially for human-AI collaboration and high-stakes domains.
- Build secure, energy-efficient hardware-software stacks that are robust to distributional shifts, adversarial attacks, and real-world degradation.