为什么机器人一到长程操作就容易掉链子？τ₀-VLA把Agent式的规划与执行分层搬进机器人。 高层策略带记忆，生成子任务，并在需要时借助世界模型搜索备选；低层用Qwen3.5视觉语言骨干和动作专家执行。仓库含训练、适配、服务与评测代码，但当前公开服务仅支持关节控制。

We also studied the scaling effect of such high-level test-time computation: more compute leads to better decisions. On out-of-distribution Book Organization, next-subtask accuracy improves from 50%…