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.
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.
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.