New preprint out on BEAR: Benchmarking and Enhancing Multimodal Language Models for Atomic Embodied Capabilities.