智元机器人发布并开源世界模型EVAC与评测基准EWMBench，助力具身世界模型加速进化！

Core Viewpoint - The article highlights the significant breakthroughs by ZhiYuan Robotics in the field of embodied intelligence, introducing the world's first action sequence-driven embodied world model EVAC and the evaluation benchmark EWMBench, both of which are now open-source. These innovations aim to establish a new development paradigm of "low-cost simulation - standardized evaluation - efficient iteration" to empower global research in embodied intelligence and accelerate technology implementation and industry development [1][21]. Group 1: Industry Challenges - The evolution of embodied intelligence faces two key constraints: high costs and risks associated with real machine validation during testing, and the lack of an efficient utilization mechanism for vast amounts of real machine data, which limits diversity generation and generalization training [3][21]. - ZhiYuan Robotics aims to address these challenges by leveraging its technical expertise and insights into industry pain points, launching the action sequence-driven world model EVAC and the evaluation benchmark EWMBench to redefine the development paradigm of embodied world models [3][21]. Group 2: Technological Breakthroughs - EVAC represents a dynamic world model capable of reproducing complex interactions between robots and their environments, marking a transition from traditional simulation to generative simulation [5][21]. - The core capabilities of EVAC include precise mapping from "physical execution" to "pixel space," enabling end-to-end generation through a multi-level action condition injection mechanism [7][21]. Group 3: Dual Value Proposition - EVAC introduces a generative simulation evaluation scheme that addresses the high costs and risks of real machine evaluations, allowing for interactive evaluation pipelines that significantly enhance the efficiency of strategy model screening [9][10]. - The data augmentation engine of EVAC can generate large-scale data from minimal expert trajectory data, leading to a task success rate increase of up to 29% for strategy models trained with this augmented data [10][21]. Group 4: Evaluation Benchmark EWMBench - EWMBench is the world's first evaluation benchmark for embodied world models, designed to fill industry gaps and establish a unified, credible evaluation standard [12][21]. - The benchmark features a three-dimensional evaluation system focusing on scene consistency, motion correctness, and semantic alignment and diversity, utilizing advanced metrics for precise assessment [15][20]. Group 5: Collaborative Synergy - The synergy between EnerVerse, EVAC, and EWMBench creates a "spiral evolution" where EnerVerse provides a robust framework for EVAC, while the diverse high-quality data generated by EVAC continuously optimizes the EnerVerse model [18][21]. - The combination of EVAC and EWMBench has been officially selected as the baseline system and evaluation standard for the AgiBot World Challenge @ IROS 2025, offering a valuable platform for developers and teams engaged in embodied intelligence research [19][21].