Submit a policy提交策略
Your policy runs read-only over recorded human operation on the bodies you pick, and you get a divergence report; it never drives a robot.你的策略在所选本体的人类操作录制上只读运行,你拿到一份偏离报告;它从不驱动机器人。
EXAMPLE Filled with a made-up policy so you can see a complete submission.预填了一个虚构策略,让你看到完整的一次提交。 i
What the report answers, and what it refuses to报告回答什么,拒绝回答什么
Answers回答
- Per body: divergence from the recorded operator, with intervals逐个本体:与录制操作员的偏离,带区间
- Whether the policy beats the two trivial floors (predict nothing; repeat the last action)策略是否胜过两条平凡下限(什么都不预测;重复上一个动作)
- Whether the cross-body change exceeds what the bodies alone already produce with no model in the loop跨本体的变化是否超过没有模型时本体差异本身就会造成的变化
Refuses拒答
- Whether the policy completes the task. Only the Cell tier, with a real robot and an operator present, can tell you that.策略能否完成任务。只有 Cell 档(真机 + 操作员在场)能回答。
- How your policy ranks against anyone else's. One policy per report, no leaderboard.你的策略和别人比排第几。一份报告一个策略,没有排行榜。
- Anything at all from a task where the human's actions barely change (step 3).人的动作几乎不变的任务,什么都不回答(见第 3 步)。
RC Shadow · submission form preview · nothing on this page is sent anywhere.RC Shadow · 提交表单预览 · 本页不会发出任何东西。