Recording Guide
Create high-quality embodied AI training videos
Videos recorded following this guide will be used directly to train embodied intelligence models. High-quality POV videos are the key data for robots to learn to operate like humans.
Why Quality Matters
Video quality directly affects model training effectiveness
Embodied AI robots need to learn operational skills from human first-person videos. Every detail of the video—hand position, motion trajectory, object interaction—will be absorbed and learned by the model.
Low-quality videos can cause the model to learn incorrect motion patterns or miss key steps, while high-quality videos significantly improve model generalization and success rates. Every careful recording you make contributes to the capability improvement of next-generation robots.
Five-Dimensional Quality Standards
Each dimension has clear qualified and excellent criteria
Hand Visibility
Practical Tips
- Keep hands in the center of the frame
- Avoid moving out of view
Stability
Practical Tips
- Use head-mounted/chest-mounted camera
- Avoid handheld shooting
Completeness
Practical Tips
- Start from picking up the tool
- Record through to completing the operation
Lighting
Practical Tips
- Face the light source while operating
- Avoid backlighting
Composition
Practical Tips
- Point camera down at the operation area
- Distance 40-70cm
Recommended Devices
Choosing the right recording device is the foundation of high-quality video
Head-mounted Camera
The most stable and natural first-person perspective
Chest-mounted
Good stability
Handheld
Excessive shaking
Recommended Parameters
Follow these settings for best results
| Parameter | Standard | Recommended |
|---|---|---|
| Resolution | 1920×1080 (minimum) | 4K (recommended) |
| Frame Rate | 30fps (standard) | 60fps (fine operations) |
| Orientation | Landscape, camera facing down | Landscape, camera facing down |
| Duration | >3 minutes (short tasks) | >10 minutes (complex tasks) |
POVHands App
One-stop recording, annotation, and upload tool
Real-time AI Detection
Use MediaPipe hand keypoint detection while recording for instant hand visibility feedback
Automatic Annotation Generation
No post-processing needed; automatically generate annotation data required for training
One-Click Upload
Upload to platforms and submit to the POVHands dataset in one click
Instant Quality Feedback
Get five-dimensional quality scores and improvement suggestions immediately after upload
Quick Checklist
Confirm the following before each recording
Once all checked, you can start recording — good luck!