Generalist

Generalist is discussed in 170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人 through Gen 1, a robotics model that the episode says improved execution speed and claimed very high success rates on complex long-horizon tasks. The source emphasizes that Generalist resists being labeled as either a world-model company or a VLA company, preferring to frame its route as training directly on physical interaction data.

E244｜端到端vs上下分层：机器人路径之争，正在转向？ adds a critique from Han Zheng / 韩正’s 速度科技 / Sudu Technology route. The episode treats Generalist’s real interaction data as important, but questions whether hundreds of thousands of hours of real data are enough for arbitrary object manipulation without Structured 3D Robot Data, Sim2Real, and lower-level skill grounding.

Key Points

• Gen 1 is described as using roughly 500,000 hours of real-world interaction data and UMI-style body-free collection.
• The source says Generalist does not simply fine-tune a pretrained VLA, but trains end to end on its own data.
• The company becomes a case for Embodied Robot Data Paradigms and for the ambiguity of model labels inside World Model VLA Fusion.
• The Sudu source makes Generalist a test case for whether Real Robot Data Strategy can scale to Open-World Robot Manipulation without a stronger simulation and structured-data layer.

Connections

• World Model VLA Fusion — broader model-label convergence the source highlights.
• Real Robot Data Strategy and Embodied Robot Data Paradigms — data route attached to Gen 1.
• Vision Language Action Models and World Models — labels the company reportedly avoids.
• 速度科技 / Sudu Technology, Sim2Real, and Layered Robot Architecture — comparison route added by E244.