Rohan Paul· @rohanpaul_ai · X·· 6 小时前AI 评分55
AI 导读
Verne 的 V0.5 在完全不进行任务特定微调的情况下，仅增加预训练数据，就把 RoboTwin 2.0 的成功率从 23% 提升到约 51%。作者指出这一趋势尚未出现放缓迹象，说明机器人操作可能存在数据规模定律，Verne 正据此把数据规模扩大 10 倍。
正文
this is cool.
feeding V0.5 more pretraining data, with no task-specific fine-tuning at all, roughly doubled its RoboTwin 2.0 success from 23% to 51%.
And the trend shows no sign of flattening. That points to a data scaling law for robot manipulation, which is why Verne is now scaling its data 10x.
来源：Rohan Paul · x.com