Cut the meat
HUMAN + HAND POSE
UR5Ewrist error – cm
Task
Subgoal
Hands
0:00 / 0:10
left handright hand
Southeast Asia · First-Person Video · Robot-Ready Data
Teach robots how people actually work.
Real people. Real environments. Real tasks.
Khora captures everyday human activity across homes, kitchens, and workplaces in Southeast Asia, then transforms each recording into robot-training-ready data.
From first-person video to hand and camera pose, layered narration, and robot-replayed motion, every recording moves through the same pipeline from capture to deployment.
This is one of our own recordings — from a human doing the task to a robot learning the motion.
Our recording · phone camera
One clip, two bodies
On the left, our tracking pipeline captures the person’s hand motion frame by frame. On the right, the same motion is retargeted to a UR5e robot arm. Below, our model generates time-aligned action captions.
Robot choice is per clip: the same recording is also available for 10 other robots.
"Wrist error" is how far the robot's wrist is from the person's at that moment.
Our recording · phone camera
Narration at three levels
Each clip includes time-aligned task, subgoal, and per-hand action labels, available in both English and Vietnamese.
Trim green beans
HUMAN + HAND POSE
Task
Subgoal
Hands
0:00 / 0:24
left handright hand
The product
What a Khora clip contains
Each clip is packaged as robot-ready training data, designed to drop directly into LeRobot or RLDS without extra processing.
Video
Egocentric video
1080p at 30 fps from head- or chest-mounted cameras, with camera intrinsics and quality scores for blur, exposure and hand visibility.
Pose
Hand and camera pose
21 keypoints per hand and a 6-DoF camera trajectory for every frame, aligned in a shared metric world frame.
Language
Three-level action labels
Time-aligned task, subgoal and per-hand Action labels, generated in both English and Vietnamese.
Robots
Robot-ready motion
Joint-space trajectories and rendered robot replays for arms and humanoids, with per-frame tracking error for quality filtering.
Scene
Scene reconstruction
Hand masks, hand-erased background plates and metric depth, supporting tracking, retargeting and downstream scene understanding.
Coverage
Real-world task diversity
Cooking, cleaning, laundry, shopping and care tasks across homes, objects, environments and languages.
Pipeline validation · public datasets
Our pipeline on public benchmarks
We run the same pipeline on public egocentric datasets to validate tracking, retargeting and narration. These examples demonstrate the pipeline, not our commercial data.
Fry an egg
HUMAN + HAND POSE
UR10E · RETARGETwrist error – cm
Task
Subgoal
Hands
0:00 / 0:15
left handright hand
Portion out snacks by scooping
HUMAN · ARIA GLASSES
Task
Subgoal
Hands
0:00 / 0:20
What is an hour of this data worth?
Help us decide what to collect first and how to price it. Two minutes, all multiple choice.
What we'll ask
- How you'd use the data
- Your organization and role
- What you spend on training data today
- Where your data comes from
- Which data components you need
- What one hour is worth to you
- How much you would need and what matters when buying