Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Server error while post-processing the rows. This occured on row 13. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

DiTaR 真机测试权重(2026-11-08)

三个任务,每个任务一个独立的包。zip 和同名文件夹的内容一样,下哪个都行。

任务 包 先测这两个
knife:把刀放到手套上 deploy_package_knife_v2 ours.pt 和 bc_noreward_aug2.pt
bottle:把水瓶放到盘子上 deploy_package_bottle_v2 ours.pt 和 bc_noreward.pt
basket:把桌上的东西放进篮子 deploy_package_basket_v2 ours.pt 和 bc_noreward.pt(一定要加 --max-steps 450)

下载

hf download EndeavourDD/ditar_re deploy_package_knife_v2.zip --repo-type dataset --local-dir .
unzip deploy_package_knife_v2.zip

bottle、basket 把文件名里的 knife 换掉就行。也可以在网页上直接下载 zip。仓库是公开的,不需要 token。

怎么用(三个包一样)

cd deploy_package_knife_v2
pip install -r requirements.txt                                        # 6 月装过就跳过
python deploy.py --ckpt checkpoints/ours.pt --mode smoke               # 不接机器人,检查能跑
python deploy.py --ckpt checkpoints/ours.pt --mode robot --episodes 1  # 上机器人跑 1 个回合
  • 第一次跑之前,务必按包里 README 的「接口」一节核对:图像是 RGB 顺序;state 是 7 个关节角(弧度)加夹爪宽度(米);action 是绝对关节目标,不要加到当前关节角上。机械臂在起始姿态时,第一步动作应该和当前关节几乎一样。如果一开始就往别处偏,多半是这几条没对上。
  • deploy.py 里机器人那部分是空壳。6 月已经接好机器人的,继续用你们自己那份 deploy.py,只把 --ckpt 指向新权重;第一次接线看包里的 README_v1_english.md。
  • 场景怎么摆、相机怎么对、怎样算成功,看每个包里的 README.md 和 scene_reference/ 里的图。
  • 测试:先测上表的两个权重,各 10 次左右,两个交替着跑。每次记成功还是失败,最好录视频。

English

Three tasks, one self-contained package per task. The zip and the folder with the same name have identical contents.

Task Package Test these two first
knife: put the knife on the glove deploy_package_knife_v2 ours.pt and bc_noreward_aug2.pt
bottle: put the water bottle on the plate deploy_package_bottle_v2 ours.pt and bc_noreward.pt
basket: put the objects on the table into the plastic basket deploy_package_basket_v2 ours.pt and bc_noreward.pt (always pass --max-steps 450)

Download (public repository, no token needed):

hf download EndeavourDD/ditar_re deploy_package_knife_v2.zip --repo-type dataset --local-dir .
unzip deploy_package_knife_v2.zip

Run (same for all three packages):

cd deploy_package_knife_v2
pip install -r requirements.txt                                        # skip if installed in June
python deploy.py --ckpt checkpoints/ours.pt --mode smoke               # check without the robot
python deploy.py --ckpt checkpoints/ours.pt --mode robot --episodes 1  # one robot episode
  • Before the first run, check the "Interface" section of the package README. The image must be in RGB order. The state is 7 joint angles in radians plus the gripper width in meters. The action is an absolute joint target, not something to add to the current angles. At the start pose, the first action should be almost equal to the current joint angles; if the arm veers off right from the start, one of these is usually wrong.
  • The robot section of deploy.py is a stub. If your robot is already wired into the June deploy.py, keep using your copy and just point --ckpt at the new checkpoints. For first-time wiring, see README_v1_english.md in each package.
  • Scene layout, camera alignment, and the success criterion are in each package's README.md and the pictures in scene_reference/.
  • Test the two checkpoints in the table first, about 10 episodes each, alternating between them. Record success or failure for every episode, ideally with video.
Downloads last month
105