GR-Dexter Technical Report (byte-dexter.github.io)

The GR-Dexter Technical Report introduces a groundbreaking framework for vision-language-action (VLA) models tailored for bimanual dexterous-hand robots. Unlike traditional systems that primarily utilize grippers, GR-Dexter employs a sophisticated, high-degrees-of-freedom (DoF) robotic hand in conjunction with an innovative teleoperation setup, allowing for the collection of diverse and scalable demonstration data. This integrated approach leverages teleoperated trajectories alongside extensive vision-language datasets, significantly enhancing the robot's ability to perform complex manipulation tasks in real-world scenarios, achieving remarkable success rates even with unseen objects and instructions. This advancement is particularly significant for the AI/ML community as it addresses key challenges in bimanual robotic manipulation, such as handling expanded action spaces and occlusions. By adopting a Mixture-of-Transformer architecture and co-training strategies using a mix of cross-embodiment and human demonstration data, GR-Dexter demonstrates robust generalization capabilities. Real-world evaluations showcase its proficiency in executing long-horizon tasks like vacuuming and bread serving, marking a significant step towards versatile generalist manipulation in robotics. The comprehensive methodology and results in this report pave the way for future developments in AI-driven robotic dexterity and adaptability.