A new arXiv paper argues that overall accuracy scores hide which stage of affordance prediction vision-language models actually fail at. The authors break the task into locating the relevant object part and knowing what action to apply, then test models on each separately. Their results indicate the bottleneck lies in part grounding rather than action knowledge.
A new arXiv paper introduces Open-UniMo, a framework that aims to combine human motion generation with motion understanding in a single model for open-world settings. The authors note that most existing motion-language models treat motion as a secondary modality attached to language, which limits how well they generalize. The work targets embodied AI systems that need to both produce and interpret human actions.
A new arXiv paper argues that collective intelligence in artificial multi-agent systems depends on how agents are organized, not just on individual capabilities. The authors note that most such systems rely on fixed organizational structures even when operating in physical environments, and propose the ORCH framework to organize embodied agents more adaptively.
A new arXiv paper introduces ReactHuman, a physics-grounded benchmark designed to evaluate how well embodied multimodal large language models handle sudden physical hazards. The tasks include scenarios such as catching a slipping plate or dodging a falling knife, which the authors frame as both a test of embodied intelligence and a prerequisite for using MLLMs as decision cores in household robots. The work is listed as a cross-submission announcement in arXiv's cs.AI category.
A new arXiv preprint argues that large-scale models still fall short when their capabilities are transferred into embodied agents. The authors propose a developmental framework that ties autonomy, social norms, and alignment together for autonomous artificial agents. The work is a conceptual research contribution rather than a system release.
A new arXiv paper examines how physical-state inputs shape a 0.8B hybrid language model adapted for robotic manipulation with only 6.2M trainable parameters. The researchers train six conditions on three LIBERO-Spatial tasks and assess robustness across three seeds and 540 held-out rollouts. The results provide training controls and diagnostic measures for geometry conditioning in small embodied models.
A new arXiv paper introduces Valerant, a system that automatically creates explorable game maps using action-conditioned world models. The approach builds on World Action Models, which combine predictive modeling with action generation so that anticipated future states can steer agent behavior. The work aims to address the limited exploration of general-purpose applications of such models in embodied AI.
Researchers introduce tinyDSM, a framework that lets centimeter-scale millirobots with very limited onboard resources autonomously explore and build skills over their operating lifetime. Drawing on developmental mechanisms and reinforcement learning, the approach targets continuous adaptation on tiny hardware.
A new arXiv research paper contends that public debate over embodied AI focuses too much on job displacement while overlooking a more fundamental danger. The authors identify governance lag—the gap between how quickly embodied AI systems are deployed in measurable ways and how fast institutions can develop the capability to respond—as the primary risk. They argue that closing this institutional timing and capacity gap is essential to managing the technology safely.
A new arXiv paper introduces VANTAGE-Bench, a benchmark that tests how well vision-language models handle infrastructure-focused video as they move toward physical deployment. The authors argue that existing evaluations center on embodied AI using subject-centric consumer footage, leaving infrastructure AI largely unexamined. The benchmark is intended to quantify this gap between current model capabilities and real-world infrastructure monitoring needs.
A new arXiv paper argues that capabilities in Physical AI systems can arise from fundamentally different origins, and that existing taxonomies based on morphology, architecture, learning algorithms, tasks, or domains do not explain where a capability actually comes from. The authors propose a framework that traces capability formation back to seven distinct sources. The goal is to give researchers a clearer way to compare how embodied systems acquire their skills.
Researchers behind the Show-Harness project propose an embodied harness that lets general-purpose vision-language models operate robots directly. The approach aims to bridge the gap between the broad world knowledge of VLMs and the practical challenge of turning that intelligence into robot control.