群核科技联手英伟达、英特尔、浙大，三篇 ECCV 论文给物理 AI 造基础设施

TL;DR - Qunhe Technology and academic and industry partners presented three ECCV 2026 works spanning simulation, self-evolving multimodal reinforcement-learning data, and spatial-agent evaluation. Together, they form an infrastructure stack for training, improving, and testing embodied AI systems.

• SPEAR exposes Unreal Engine through Python and uses zero-copy transfer and asynchronous execution, achieving 73 FPS at 1080p and substantially higher rendering throughput than UnrealCV+ and AirSim.
• Syn-GRPO asynchronously generates increasingly challenging, annotation-preserving visual scenes during training to counter entropy and diversity collapse, adding about 5% training overhead.
• WalkerBench evaluates vision-only agents through interactive navigation across more than 160 cities; its Spatial-IDE framework externalizes topological memory and separates high-level reasoning from local perception.
• Spatial-IDE improved nine tested VLM agents by 104.97% on average, and was also deployed on a Unitree G1 for kilometer-scale urban navigation without environment-specific fine-tuning.