Choose theme
AI Resources
UnifoLM-WBT-Dataset
UnifoLM-WBT-Dataset is Unitree's collection of recorded G1 whole-body teleoperation tasks, with video, robot-state and action data.
Choose an individual task and hand configuration from the collection, then open its dataset card and files. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
Robotics dataset collection
Unitree publishes separate demonstration datasets for its G1 humanoid, including household manipulation tasks.
Why it stands out
Data behind the task videos
The linked cards describe both camera recordings and numerical observations and commands from the demonstrations.
Availability
Hugging Face collection from Unitree
Hugging Face hosts the collection, individual dataset files, episode previews and schema descriptions.
Why it matters
What makes it useful
For someone studying robot demonstrations, the dishwasher card provides head and wrist camera streams alongside recorded observations and actions. Its fields include arm and leg joints, hand states, target hand commands and timestamps. A researcher can examine the recorded commands as well as the video, rather than only watch a robot complete a task.
What to know
Where it fits
These are demonstration datasets, not a trained controller or a robot application. Even the same activity has separate hardware entries: the washing-machine task appears with Inspire and Dex1 hands. The task name gets a reader to the activity; the hand label identifies which recording setup the entry describes.
Notable points
What stands out
The linked cards describe whole-body signals as well as manipulation. Their lower_body fields include both legs and the waist, while base_command records forward and lateral velocity, yaw angle and height. The dishwasher entry lists four camera streams: a head stereo pair and one camera at each wrist. These details show what a data loader needs to handle beyond arm movements alone.
Before using
What to review
Open the chosen task's own card and files: collection membership does not mean every entry has the same episode count or hardware configuration.
The linked cards identify codebase v3.0, Parquet data and AV1 camera recordings at 30 fps. Match the loader and video decoder to those formats.
End-effector poses are supplied in base and torso frames. Select the coordinate frame described by the field rather than treating the two pose fields as interchangeable.
Read the selected dataset's licence and collection notes. The three linked task cards name Apache-2.0; that label describes these data entries, not a robot-use approval.
Decide whether recordings from the selected setup fit your robot, operator, objects and environment before using them for training or transfer. Hardware compatibility and human-safety review are separate decisions; a dataset or licence does not approve physical operation.
Reader fit
Who may find it relevant
Researchers examining teleoperated whole-body and manipulation demonstrations.
Robotics data teams building loaders for joint, camera, hand and coordinate-frame fields.
A reader looking for a trained robot policy will need a different artifact; this collection supplies recordings and their data descriptions.
Editorial note
Why LifeHubber lists it
UnifoLM-WBT-Dataset earns a place for data teams comparing manipulation recordings from different hands: its task cards explain the finger and channel meanings needed to interpret those recordings. Before combining hand data from different entries, use the card's Hand Convention table to map channels to fingers. Inspire lists the index finger first; BrainCo lists the thumb's open/close channel first, although both describe six channels per hand on a 0-to-1 open-to-close range. Dex1 instead lists an open/close range from 5.5 to 0. Matching an array's shape alone would miss those differences in order and meaning. This supports comparison and loader preparation; it does not establish that a policy trained on one setup will transfer to another robot.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in Embodied / Physical AI
Keep browsing this category
Explore more embodied AI resources.
LeRobot
huggingface/lerobot
A Hugging Face robotics library for robot-learning workflows, with models, datasets, robot interfaces, training, evaluation, deployment, and v0.6.0 updates across world models, reward models, benchmarks, dataset tooling, and cloud training paths.
LeRobot Humanoid
Virgileboat/lerobot-humanoid
A public humanoid robotics workspace around a low-cost 3D-printed robot-learning platform, with hardware files, assembly documentation, runtime tools, simulation/model assets, identification tooling, and training environments.
LabClaw
wu-yc/LabClaw
Scientific agent skills for analysis, literature research and visualization, including diagram generation with specified labels and relationships through an external image service.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
