ShengShu Technology has launched Vidar, a new system using synthetic data to solve the industry’s training data bottleneck, making AI robot training more efficient, scalable, and affordable for real-world use.
A significant hurdle in developing humanoid robots is the immense amount of training data required. This data collection process is expensive and time-consuming, a challenge that has traditionally slowed progress in the robotics field.
How ShengShu Technology Vidar Works
ShengShu Technology introduced its Vidar system to address the robot training data bottleneck. Vidar, which stands for Video Diffusion for Action Reasoning, generates synthetic training environments from a small amount of real video.
This method blends real-world footage with AI-generated video, utilizing the company’s Vidu video model. This approach creates a more efficient and scalable training process than traditional methods requiring physical interaction.
A New Approach to AI Robot Training
The system works by decoupling perception from control, using a task-agnostic system called AnyPos to translate learned knowledge into motor commands. Vidar can simulate complex, lifelike scenarios in a virtual space, a key departure from older training techniques.
This use of synthetic data for robot training is remarkably efficient, requiring only about 20 minutes of initial data. According to CyberGuy.com, this is between 1/80 and 1/1200 of the data required by leading models.
Unlocking Future Humanoid Robot Applications
Vidar’s design means robots can adapt more quickly to new tasks and environments. This breakthrough could unlock numerous real-world humanoid robot applications, including eldercare, home assistance, healthcare, and smart manufacturing.
By solving key issues of cost and scalability, the technology brings the concept of a household robot helper closer to reality. This scalable training could significantly speed up the deployment of practical robots in everyday settings.
