DexMimicGen: Generate 1,000 New Demos with Just 5 Demonstrations

Unlock the future of robotics with DexMimicGen! Learn how just five demonstrations can generate 1,000 new robot demos, achieving up to 97% success in simulations.

Humans only need to demonstrate a complex skill five times for a robot to learn it.

NVIDIA Labs has proposed a new solution to the lack of training data for robots—DexMimicGen.

After five demonstrations, DexMimicGen can directly mimic 1,000 new demos.

Its usability is strong; robots trained with these new demos can achieve a task success rate of up to 97% in a simulation environment, outperforming those trained with real human data.

Jim Fan, a scientist at NVIDIA involved in this project, believes that this machine-to-machine training approach addresses the biggest pain point in robotics—data collection.

Additionally, Jim Fan predicts:

"The future of robot data is generative, and the entire robot learning process will also be generative."