Evaluating embodied intelligence remains a fundamental challenge in the field. Physical-world testing often fails to control all variables, resulting in unreproducible outcomes, high replication costs, and difficulties in conducting large-scale assessments. While simulation-based evaluation addresses reproducibility issues, current virtual platforms suffer from significant reality gaps and lack proper benchmarking against real-world tests, compromising result credibility.

As a core part of the Workshop on Multimodal Robot Learning in Physical Worlds, IROS 2025, this competition considers the above challenges and features two common embodied tasks across dual phases: a simulated round hosted on InternUtopia, a high-fidelity simulation platform, followed by real-world testing. This dual-phase design aims to drive innovation in both model architectures and training methodologies within the field.

Participants are required to utilize interactively collected data from either real-world or simulated environments to accurately interpret complex scenarios and make contextually appropriate decisions. The challenge explores key issues in transferring skills from simulation to reality, including domain gaps, multimodal information fusion (vision, language, and action), and scalable sim-to-real transfer techniques.

We hope this event will bring together researchers and practitioners from around the world to explore cutting-edge topics in multimodal robot learning, laying a solid foundation for the future development of intelligent robotics.