Submit for Evaluation提交方案评估

1 · Run a reproducible evaluation1 · 完成可复现的评估

1. Check robot calibration, follower H poses, camera order and reference overlays using the calibration guide ↗. Record the node ID and any hardware differences.按系统标定指南 ↗检查机械臂校准、从臂 H 位、相机顺序与参考图叠加，记录设备编号和硬件差异。
2. Choose the task and instruction from the task reference ↗. Record the training datasets, checkpoint revision and scene references. Website task numbers and repository directory numbers differ for some tasks; include both.按任务参考 ↗确定任务与指令，记录训练数据、权重版本和场景参考图。部分任务的网页编号与仓库目录编号不同，请同时注明。
3. Run ID and OOD separately with the same checkpoint. The reference set currently documented in the guide provides ID images; OOD runs need their own defined reference images and a description of what differs from training.使用同一权重分别评估 ID 与 OOD。指南当前收录的参考集提供 ID 图；开展 OOD 评估前，需准备对应参考图并说明与训练条件的差异。
4. Use the existing commands in the evaluation guide ↗ for ACT, SmolVLA, XVLA, RDT or OpenPI. Validate inputs and motion using the documented dry-run or limited-motion checks before the real run.使用运行评估指南 ↗中 ACT、SmolVLA、XVLA、RDT 或 OpenPI 的现有命令。先按指南完成 dry-run 或最小动作检查，再进行正式实机评估。
5. Reset the scene against the selected reference before each trial. Save every scored trial, including failures, timeouts and interventions, with its outcome and video. Dry-run checks are not evaluation results.每回合按所选参考图复位。保留所有计分回合的结果与视频，包括失败、超时和人工干预；dry-run 检查不计为评估成绩。

RDT and OpenPI use separate clients with different example defaults. Report their actual episode count, time limit and inference settings; agree on matching conditions before comparing scores. The table summarizes the current LeRobot template.RDT 和 OpenPI 使用独立客户端，示例默认值有所不同。请报告实际回合数、时限与推理设置，比较成绩前统一条件；上表说明现有 LeRobot 模板。

2 · Prepare the submission materials2 · 整理提交材料

• Method: team or anonymous team ID, method name, code commit, checkpoint version, training dataset IDs and revisions, pretraining and additional data.方案说明：团队或匿名团队编号、方案名、代码提交版本、权重版本、训练数据集 ID 与版本、预训练及额外数据。
• Setup: node IDs, hardware, calibration checks, camera references and settings, task instructions, ID/OOD definitions and differences from the standard setup.设备与条件：设备编号、硬件、校准检查、相机参考图与参数、任务指令、ID/OOD 定义及与标准配置的差异。
• Reproduction: the exact commands and configuration files used, software versions, episode count, time limits, seeds, inference settings and success criteria.复现配置：实际运行命令、配置文件、软件版本、回合数、时限、seed、推理设置与成功判据。
• Evidence: a per-trial result table, evaluation videos, logs and a summary by task, node and condition. Link large files through an accessible repository or file host.结果证据：逐回合结果表、评估视频、日志，以及按任务、设备和条件汇总的结果。较大文件请提供可访问的托管链接。

3 · Report results by task and condition3 · 按任务和条件报告结果

For each task × node × ID/OOD condition, report successful trials, total scored trials and success rate. Success rate = successful trials / scored trials × 100%. Define the success criterion and list any excluded setup errors with reasons.每个任务 × 设备 × ID/OOD 条件分别报告成功数、计分回合总数和成功率。成功率 = 成功回合数 / 计分回合数 × 100%。注明成功判据；因设置错误排除的回合需单独列出原因。

Keep ID and OOD results separate. Report each node before aggregating across nodes. State task coverage and the averaging method; mark untested tasks as untested. Submit complete trial records, not only successful clips.ID 与 OOD 分开报告，多设备先分别报告再汇总。注明已评估任务范围和平均方式，未评估任务明确标注；提交完整回合记录，不只提供成功片段。

4 · Submit and follow up4 · 提交与后续沟通

Maintainers review the setup, trial records and evidence before updating the leaderboard. If information is missing, follow up in the same Issue or PR. Current leaderboard entries are illustrative while the protocol and official results are finalized.维护者核对配置、逐回合记录与证据后更新排行榜；材料需补充时，在同一 Issue 或 PR 中继续跟进。当前榜单条目仍为规程与正式成绩定稿前的示意。