Solutions · Humanoid robot data
Humanoid robot data
Humanoid robot data is demonstration data of a two-armed embodiment doing real tasks — video from head and wrist cameras paired with joint states — used to train the manipulation policies and foundation models that run humanoids.
PrismaX collects it by teleoperating robot arms, including its own bimanual YAM rigs with an overhead camera and a camera on each wrist, and by recording first-person video of people doing hands-on work.
The data problem
What gets in the way
Stuck in the lab
The embodiment gap
Teleop costs time
Hands before legs
What the data needs
A humanoid-ish view of real work.
- A head or overhead camera with full view of the workspace, plus a camera on each wrist, as in our Standards for Data.
- Bimanual embodiments with six or seven degree-of-freedom arms and 1-DoF grippers or dexterous hands.
- Joint states for every joint and a matching URDF, so the configuration can be reconstructed and transferred across embodiments.
- Whole tasks completed with smooth, mistake-free motion.
How PrismaX provides it
From the catalog, or collected to spec.
Teleoperation data
Egocentric data
Custom data collection
Industries and scenes
Where the demonstrations come from
- HomeTidying, sorting, folding, placing
- KitchenFood handling and clean-up
- LogisticsPacking, unpacking, bin picking
- IndustrialComponent handling
Related tasks
- Pick and place various fruits into containersKitchen · 2.8k
- Pick and place flowers into vase and vase on shelfHome · 2.0k
- Block manipulationHome · 2.0k
- Pack box of componentsLogistics · 1.8k
- Pick and place slippers into binHome · 1.6k
- Tying shoelacesHome · 1.5k
- Clean up kitchen areaKitchen · 1.4k
- Unpack box of objects into binsLogistics · 1.2k
Sample videos
See the episodes
Two robot arms arrange flowers in a vase, seen from the overhead and wrist cameras at the same time.
A montage of teleoperated sessions folding T-shirts, seen from many angles.
Related reading
All news
Deep Dive
GR00T Part 1: Foundation Models & the History of AI Robotics
Nvidia’s GR00T-1 marks a GPT2 moment for robotics—proof that foundation-scale models trained on video can generalize across tasks and hardware.
Deep Dive
GR00T Part 4: Advancing Robotics Foundation Models
In this final part, we explore how robotics foundation models can move from GPT-2 to GPT-3 era using synthetic, simulator, and high-quality data.
Deep Dive
Teleoperation: Now and the Future of AI Robotics
Read this article if you’re interested in contributing to the robotics revolution, which will reshape the way we work and interact with the physical world.
FAQ
Frequently asked questions
What is humanoid robot data?
Demonstrations of a two-armed, human-like embodiment completing real tasks: video from head and wrist cameras paired with the angle of every joint. Humanoid manipulation policies and foundation models learn from it.
Do you collect data on humanoids or on robot arms?
Today the catalog is robot-arm data on YAM, Piper, RLM 63 and TOK2, including bimanual YAM rigs with an overhead camera and a camera on each wrist. For other embodiments, describe what you need and we scope a custom collection around it.
Is human egocentric video useful for humanoids?
It is cheap to capture and the motion is smooth, but there is a cross-embodiment gap: some human motions are well outside what the hardware can do. As practitioners we prefer teleoperated data, which is usually higher quality and easier to train on, and offer egocentric video alongside it.
What sensor data comes with each episode?
Synchronized overhead and wrist camera video and the angle of every joint, with a URDF that matches the kinematic chain.
Can I start without a custom order?
Yes. Discover and Researcher give self-serve monthly access to the catalog. Enterprise adds custom collection to your spec.
Next step
Data for the robot you’re building.
Tell us the robot, the tasks and the scenes. We'll match it to the catalog or scope a collection.