具身智能：长上下文世界模型推动端到端控制，但操控仍是瓶颈
Key Questions
What new world model did NVIDIA release?
NVIDIA Cosmos 3 introduces a full-modality world model with Flash-WAM delivering 23x acceleration for robotic applications.
Which open-source VLA model was released?
Alpamayo 2 Super 32B VLA has been open-sourced to advance vision-language-action capabilities in robotics.
What are the main bottlenecks in embodied AI?
Fine-grained manipulation, data loop closure, and overall system cost remain the primary challenges despite progress in world models.
How are humanoid robots becoming more accessible?
Unitree R1 humanoid robots now start at $4,900, crossing the gaming-PC price threshold while supporting walking and basic tasks.
What research advances spatial reasoning for robots?
SpatialClaw provides a spatial reasoning action interface, and LabVLA translates laboratory protocols into robot execution.
Cosmos 3、Alpamayo 2、Flash-WAM与Long-WAM推动视觉历史、世界模型和动作策略融合；Long-WAM特别显示长上下文需要自回归视频预训练支撑，且已考虑RTX 5090和Jetson实时部署。机器人数据闭环、精细操控、跨平台泛化和真实环境稳定性仍决定商业化速度。