Embodied AI and physical reasoning

Transferring the capabilities of large foundation models to physical robots remains a major gap. Sim-to-real transfer is fragile, and language-conditioned robot policies struggle with dexterous manipulation, contact-rich tasks, and novel environments. World models that capture physical dynamics with sufficient fidelity for planning are nascent. Building embodied agents that combine the common sense of language models with the sensorimotor precision required for real-world manipulation is largely unsolved.