DexMimicGen is an automated data generation system jointly proposed by NVIDIA Research, UT Austin, and UC San Diego in 2024. It can generate a large amount of robot training data through a small amount of human demonstrations.DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning".

The core function of DexMimicGen is to use a small amount of human-demonstrated data to generate a large number of similar demonstrations through transformation and imitation techniques. This technology has demonstrated excellent effectiveness, with a mission success rate of up to 97% in a simulated environment. The system can generate up to 1,000 robot training demonstrations with only 5 human demonstrations, solving the data scarcity problem in robot training and demonstrating the great potential of generative learning in robot training.

The researchers conducted 60 demonstration experiments in 9 scenarios, generating a total of 21,000 data demonstrations. The robots trained with the data generated by DexMimicGen achieved success rates of 76% and 80.7% in tasks such as tidying drawers and assembling building blocks, respectively, while these success rates were only 0.7% and 3.3% when trained with traditional human data.