AI 导读
🧵 3. 双先验预训练加入 2 个先验：用于重复动作模式的 Latent Action，以及用于空间结构的 Geometry。 两者都在机器人特定适配之前学习，为模型提供更强的跨具身迁移基础。https://t.co/8uV70mGGyD
正文
🧵 3. Dual-prior pre-training adds 2 priors: Latent Action for recurring action patterns and Geometry for spatial structure.
Both are learned before robot-specific adaptation, giving the model a stronger base for cross-embodiment transfer. https://t.co/8uV70mGGyD
来源：@rohanpaul_ai · x.com