July 8, 2026 · Wednesday2026 年 7 月 8 日 · No. 4 NEWSACCESSABOUT

AUTOSIGNAL 车智信号

Human-curated. Expert-annotated. Every signal traced to source. 人工精选 · 专家点评 · 每条信号可溯源

← All signals← 返回全部信号

Research Frontier研究前沿

RoboWorld Neural Simulator Reaches 98.9% Real-World Correlation in Robot Policy TestingRoboWorld神经模拟器精度达98.9%,机器人策略评估无需真机部署

S 1.7 T1 1 sources1 个来源 R7-research
  1. Researchers led by Byeongguk Jeon introduce RoboWorld, an automated evaluation pipeline for generalist robot policies using neural video world models, submitted to ICML 2026.
  2. The system achieves 0.989 Pearson correlation and 0.970 Spearman correlation with real-world robot evaluation, pairing a fast autoregressive video world model with a task-progress-aware vision-language model for policy scoring.
  3. RoboWorld proposes Step Forcing, a technique combining anchored and one-step self-forwarded contexts to reduce train-test mismatch, enabling reliable long-horizon autoregressive rollouts while maintaining fast inference speed.
  1. 由Byeongguk Jeon等研究人员领导的团队推出RoboWorld,一个基于神经视频世界模型的自动化机器人策略评估管道,已提交至ICML 2026。
  2. 该系统与真实机器人测试的相关性达到0.989(皮尔逊相关系数)和0.970(斯皮尔曼相关系数),采用快速自回归视频世界模型与任务进度感知视觉语言模型评分相结合。
  3. RoboWorld提出Step Forcing技术,结合锚点固定和单步自前进上下文以减少训练-测试偏差,在保持快速推理的同时实现可靠的长时域自回归展开。