Hydra-0: Action Flow for Generalist World Modeling and Control
- 类型：arxiv
- 标识：2608.18077
- 链接：https://arxiv.org/abs/2608.18077
- 主分类：multimodal
- 形态：benchmark
- TLDR：We introduce Hydra-0, a generalist world model conditioned on action flow, which represents robot actions as pixel motion. This shared visual interface enables generalist world modeling and control by learning action consequences across embodiments, tasks, environments, and video-generation backbones. Our best configuration achieves 90.4% lower robot-motion error and 60.2% lower object-motion error than our action-conditioned baseline, while supporting zero-shot composition and data-efficient adaptation. On the RoboLab benchmark, Hydra-0 achieves a Pearson correlation of r=0.96 between replaye
- 待LLM分类：否
- 标题中文：Hydra-0：面向通用世界建模与控制的动作流
- TLDR中文：我们提出 Hydra-0，一种以动作流为条件的通用世界模型，将机器人动作表示为像素运动。这种共享的视觉接口使得跨具身、任务、环境和视频生成 backbone 的通用世界建模与控制成为可能，学习动作在不同场景下的后果。我们的最佳配置相比动作条件 baseline，机器人运动误差降低 90.4%，物体运动误差降低 60.2%，同时支持零样本组合与数据高效适配。在 RoboLab 基准上，Hydra-0 在 replaye
- 来源文件：
- /inbox/tom/_candidates/2026-08-25-agent-rag-longcontext-candidates.json
- /inbox/tom/_candidates/2026-08-25-agent-memory-tool-use-candidates.json