Overview 总体介绍
This track focuses on real-world physical robot manipulation and adopts a multi-task strategy. In this competition, you will use your model to operate real robots to complete 30 newly designed desktop manipulation tasks (20 basic tasks and 10 bonus tasks). Your model should be able to complete any of these different tasks based on prompt instructions. There are no restrictions on the model type; both World Models and VLA models are welcome.
本赛道专注于真实物理世界的机器人操作，并采用多任务（multi-tasks）赛制策略。在本次比赛中，你将通过你的模型操作真实的机器人完成 30 个全新设计的桌面操作任务，包括 20 个基础任务和 10 个附加任务。你的模型应当能够根据 prompt 的指示完成多个不同任务中的某一个。赛道对参赛模型类型不做限制，世界模型（World Model）与 VLA 模型均可参赛。
Competition Rules & Scoring 比赛规则与计分方式
1. Resources & Quotas 1. 资源与测试配额
- Each team will receive demo data for all 30 tasks.
- Each team has a total quota of 100 + 10 task evaluation submissions (100 for basic tasks, 10 for bonus tasks).
- Regardless of whether a test is scored, every task submitted counts towards your quota (e.g., submitting a batch of 5 tasks consumes 5 from your quota).
- 每支参赛队伍将得到 30 个任务全部的演示数据。
- 每队拥有 100 + 10 次任务提交测评的配额（基础任务 100 次，附加任务 10 次）。
- 无论是否计分，每次提交进行的所有任务都会计入总测试次数的配额（如：提交 5 个任务，就会占配额中的 5 次）。
2. Testing Procedure & "Multi-task" Requirement 2. 测试流程与“多任务”要求
- Each model runs 10 consecutive times per task. On-site testers will record the success rate and step score.
- Scored Submissions: To be eligible for final scoring, you must submit in a "multi-task" format. This means simultaneously submitting all basic tasks or all bonus tasks for the same robot type using the same model. We will initiate task tests in an unannounced order.
- Unscored Submissions: You can submit any subset of tasks for testing, but if the submission does not include the full set of basic or bonus tasks for a robot type, it will be marked as "unscored" (though it still consumes quota).
- For all "bonus tasks", you only have one chance to submit. We recommend training models that can cover more tasks and submitting for testing after debugging the basic tasks.
- 每个模型在每个任务上进行 10 次连续的运行，并由现场测试员统计成功率和步骤分。
- 计分测试：如果希望进入最终计分，需要用多任务（multi-tasks）形式提交，即同时提交至少是同一机型的全部基础任务或全部附加任务（需使用同一模型）。我们会以事先不通知的顺序依次发起测试并统计总分。
- 不计分测试：你可以以任意的任务子集发起测试，但如果不包含至少一个机型的全部基础或附加任务，将成为“不计分”测试（仍消耗配额）。
- 对于所有的“附加任务”，你都只有一次测试机会。推荐你尽可能训练能够涵盖更多任务的模型，并在基础任务效果调试好之后进行测试。
3. Ranking & Prizes 3. 排名与奖励
- We will record the best submission for each robot type in both basic and bonus sets (i.e., sum of different tasks within the same robot type in a single test, and the highest sum across multiple tests).
- Total average success rate and step scores are calculated using weights: Weight 1.0 for basic tasks, Weight 0.25 for bonus tasks.
- Ranking is based on total success rate. If tied, the step score will be used as a tie-breaker.
- The Top 3 teams will receive prize money. Teams with non-zero scores can apply for a certificate.
- 我们将在基础集合和附加集合中每种机型记录最好的提交（即单次测试内同一机型在任务集合内不同任务求和，在多次测试中以总和最高的测试为准）。
- 总平均成功率和总平均步骤分将加权计算：基础任务权重为 1.0，附加任务权重为 0.25。
- 最终按总成功率进行排名，如果成功率相同，按步骤分排名。
- 比赛将角逐出前 3 名获得奖金。成绩非零的队伍都可以申请获得证书。
Notice: Table30 for CVPR Dataset Released 重要通知：Table30 for CVPR 数据集已发布
The Table30 for CVPR dataset has been officially released. You can download it from Hugging Face. The dedicated track leaderboard will be available at robochallenge.cn/competition and will be updated soon.
Table30 for CVPR 数据集已正式发布，可从 Hugging Face 下载。专属赛道榜单地址为 robochallenge.cn/competition，即将更新。
Step-by-step Guidance 参赛步骤
Step 1. Registration 步骤 1：账号注册
⚠️ Note: Each participating team is limited to registering only one account. ⚠️ 注意：每个参赛队仅限注册一个账号。
- Please use the following template to send an email to join-robochallenge@robochallenge.ai to apply for your testing Token:
- 请使用以下模板向 join-robochallenge@robochallenge.ai 发送邮件，申请测试 Token：
- After registering, check your confirmation email and scan the QR code to add our Event Manager on WeChat. You can join the official communication group to receive competition updates and exclusive technical support.
- 注册成功后，请查看确认邮件并扫描其中的二维码，添加运营负责人微信。通过后即可加入官方交流群，获取赛事资讯与专属技术支持。
Step 2. Environment & Local Mock Testing 步骤 2：环境搭建与本地 Mock 调试
Before real-robot evaluation, you must adapt your model using the official inference repository.
在进行真机测试前，需基于官方推理仓库完成环境搭建与代码适配。
git clone https://github.com/RoboChallenge/RoboChallengeInference.git
cd RoboChallengeInference
git checkout cvpr
pip install -r requirements.txt
Edit mock_settings.py to configure your target robot tag (e.g., 'aloha', 'arx5', 'ur5', 'w1'). Then run:
编辑 mock_settings.py 设置目标机器人（如 'aloha', 'arx5', 'ur5', 'w1'）。随后运行：
python3 mock_robot_server.py
Run test.py to verify that your custom policy properly interacts with the Mock interface.
运行 test.py 验证推理逻辑能否成功生成动作指令并下发至 Mock 机器人。
Step 3. Online Submission & Real Robot Execution 步骤 3：线上提交与真机执行
- Submit your test request via the RoboChallenge platform.
- The real-robot testing time will be based on your scheduled appointment, and our operations staff will communicate with you in advance.
- After submission, get your SubmissionId. Prepare your model and code before the scheduled slot.
- During your slot, use the API to get current machine info and run the formal script locally to interact with the physical device and submit trajectories:
- 通过 RoboChallenge 平台提交测试请求。
- 真机测试时间将以预约的时间为准，同时也会有运营人员提前进行沟通。
- 提交后获取你的 SubmissionId。你需要在测试开始前准备好你的模型和程序。
- 在指定时间段内，通过 API 获取当前机器信息，并在本地运行正式推理脚本，程序将与真实物理设备交互并提交动作轨迹：
python3 demo.py --user_token {your_user_token} --submission_id {your_submission_id} --checkpoint {your_checkpoint}
All timestamps below are in UTC. 以下时间均为 UTC（全球统一时区）。