FAQs

赛事流程

Q: I would like to sign up for this competition, how would I sign up? How much does the signup cost?

A: The signup process is free of charge. Go to Alibaba Cloud - Tianchi Competition - Engineering Development Competition - Leju Robotics REAL; Click Sign Up, fill in your personal information to sign up. (Note: For individual signups, school/company name can be left blank, Any teams in the signup information are not to be recorded as competition teams. For teaming up for the competition, please do so after everyone’s signup have been completed)

Q: How do I create/join teams?

A：完成报名后，在比赛页面左侧点击我的团队选项，可以创建队伍/加入队伍，每个队推荐1-4人。

Q: How do I participate in the real-machine competition? Is this an in-person only event?

A: All participants/teams passing the first qualification round (simulator round) may participate in the real-machine competition. If for any reason one is unable to partake in-person for testing, we can arrange technical personnel live on-stage to configure for your testing

Q: How to apply for GPU compute power assistance?

A: Relevant assistance resources are still under negotiation for now. We will provide an update as soon as any becomes available. We strongly encourage everyone to equip and utilise their own compute power for now.

Q: What data formats are provided for the datasets

A: Everything uploaded are in native rosbag formats, but participants can freely use python kuavo_data/CvtRosbag2Lerobot.py to freely convert such datasets into Lerobot Parquet format en masse. For more details, please consult the kuavo_data_challenge readme documentation.

Q: Does the dataset needs to be downloaded all at once?

A: We recommend you download only what you’ll need during each download. There is no need for you to download everything at once, as such download size is humongous!

Q：如何提交呢？提交后多久会有成绩？

A：提交需要将模型与参数等打包成docker镜像，具体请查看“提交说明”栏目，最后提交的是zip压缩包。提交后后台即会开始评测，评测完成后，若评测成功会上传分数，若评测失败会上传失败原因，此时请检查自己本地是否可以加载镜像并正常推理，若本地一切正常，请在群里联系管理员。

Q：提交的作品为什么一直没有被评测呢？

A：如果您的提交结果一直没有显示在评测中，说明您的作品还没有被评测人员拉取，请您耐心等待，通常1-3个工作日内会被评测，或可以联系管理员。

基准代码/环境常见问题

Q: After opening the simulator (i.e. deploy.py), the simulator opens and immediately disappears.

A: Please ensure that ROBOT_VERSION is set to 45. You can check this by echo $ROBOT_VERSION and see if 45 is correctly printed.

  • Note that after every re-entry into the docker image, you need to rerun export ROBOT_VERSION=45, or you can optionally add this line into the .zshrc file.

Q: During simutaneous execution of deploy.py and eval_kuavo.sh , the simulator opens successfully, but either the evaluation never begins, never restarts after the first reset round, and/or every round becomes a reset round in an infinite loop

A: In the running eval_kuavo.sh, check the following items in order:

  • Ensure that Option 8 was the one selected

  • Press L to check the log, ensure that there are no crash messages, potentially indicating missing Python packages (commonly apriltag_ros may be missing). Install any missing packages as prompted

  • Ensure that the pretrained weights filepath is correctly set, and that the code has sufficient permissions to access it

  • It is highly recommended that you change and grant the permissions of all folders in use with this project. It is not recommended for you to add sudo in front of eval_kuavo.sh as it may lead to unexpected errors

  • Reconfigure the Python environment containing the eval_kuavo.sh

  • Under some slower systems, it may be necessary for you to wait longer for the simulator to be ready. Change the time.sleep at approx. Line 348 to a larger value.

Q: When catkin build ing the simulator, it fails to find necessary modules, such as humanoid_interface

A: Check the following checklist:

  • Please ensure that source installed/setup.zsh successfully executed prior to catkin build

  • Do NOT use catkin clean, as it may erase critical packages. If you accidentally used this, please re-pull the entire repository to start over.

  • re-pull the entire repository

Q: When catkin build ing the simulator, errors pop up that read failed to make symbolic link .../../*.so

A: If you used git clone inside Windows environment, where symbolic links do not work correctly under Linux, this can occur

  • Please use a Linux distro to perform git clone/pull

Q: The simulator is very laggy

A: Check the following checklist:

  • Please ensure that run_with_gpu.sh was used to create the docker container, and that it is using CUDA properly.

  • Ensure that docker is correctly using CUDA

  • It is not recommended for you to use WSL (Windows Subsystem for Linux) for this project, as it is reported that it has CUDA acceleration issues. If you are under such environment, please see the QQ group chat to see if there are any existing solutions inside the discussion group.

Q: No simulator window showing up after executing deploy.py for the simulator

A: Please ensure that there are no running ROS processes that are not killed properly. Restart the computer and try again

真机部署常见问题

Q：为什么镜像在真机部署时 SDK 或接口报错？

A：常见原因是使用了仿真赛阶段的旧版 KDC 代码，或没有基于真机赛官网提供的最新代码构建镜像。请使用 真机赛文档 中的最新仓库和部署说明重新构建，并确认镜像内 kuavo-humanoid-sdk 版本符合要求。

Q：为什么容器里运行模型时提示缺少 Python 包？

A：通常是本地环境额外安装了依赖，但打包 Docker 镜像时没有带进去。建议先打包自己的 conda 环境，再按官网 Dockerfile 构建镜像：

conda install -c conda-forge conda-pack
conda activate kdc
conda pack -n kdc -o myenv.tar.gz

构建镜像前，请确认推理所需依赖都已安装在该环境中。

Q：为什么部署推理时提示缺少 ``pyaudio`` 或相关音频依赖？

A：KDC 在部署推理时通常需要 pyaudio 包，否则运行时容易报错。建议提前在 conda 环境中执行以下指令完成环境适配：

sudo apt update
sudo apt install build-essential python3-dev portaudio19-dev
pip install pyaudio
Q：为什么 ``run_with_gpu.sh`` 启动后无法连接 ROS？

A：常见原因是启动脚本没有按真机部署通知更新，尤其是 ROS 网络配置写错。请确认 run_with_gpu.sh 的 docker run 中包含：

-e ROS_MASTER_URI=http://kuavo_master:11311
-e ROS_IP=192.168.26.10
Q：为什么配置了模型，但部署时找不到权重？

A：通常是 configs/deploy/kuavo_env.yaml 中的路径信息没有填对。请检查 inference 字段是否能对应到镜像内模型路径。

inference:
  task: "your_task"
  method: "your_method"
  timestamp: "run_xxxxxxxx_xxxxxx"
  epoch: best

对应路径格式：

/root/kuavo_data_challenge/outputs/train/<task>/<method>/<timestamp>/epoch<epoch>

例如 task: "small"、method: "act"、timestamp: "run_20260429_002926"、epoch: best 对应：

/root/kuavo_data_challenge/outputs/train/small/act/run_20260429_002926/epochbest

如果使用 pretrained_path，请确认该路径在容器内真实存在。

Q：为什么末端执行器相关配置会导致部署异常？

A：请确认 obs_key_map 中的末端部分与当前配置的 eef_type 类型保持一致，其余末端执行器话题建议暂时注释，避免加载到错误的末端状态话题。

Q：为什么模型首次推理时尝试联网下载文件？

A：有些模型会在首次运行时下载缓存文件，例如 ACT 可能需要 ResNet18 相关缓存。如果比赛现场无网络，推理会失败。请提前在镜像中准备好所有缓存文件，或在构建镜像前完成一次模型加载测试，确保缓存已写入镜像。

Q：如何提交前快速自检？

A：下载自检脚本并放在镜像 tar 和 run_with_gpu.sh 同级目录下：

下载 check_docker_python_deps.sh

./check_docker_python_deps.sh your_image.tar

脚本会检查：

  • run_with_gpu.sh 中的 ROS 配置

  • Docker 镜像内 Python 依赖版本

  • kuavo_env.yaml 是否能对应到 outputs/train/... 模型路径