Unitree Technology recently announced the open-source release of its flagship product, the G1 robot's operational dataset, which has garnered significant attention in the industry. This dataset includes data collection methods, learning algorithms, training data, and related models, injecting new vitality into the field of robotics research.
It is noteworthy that Unitree chose to conduct training and testing based on the well-known LeRobot open-source framework. This decision received positive feedback from the CEO of Hugging Face, who stated that it reflects the increasing importance of open-source technology in the field of robotics.
Currently, the dataset includes five categories of basic operational capabilities, such as unscrewing bottle caps, pouring water, stacking three-color building blocks, packaging cameras, storing items, and dual-arm grasping and placing. In the data collection phase, Unitree innovatively utilized Apple's Vision Pro for remote operation control of the G1, and the related code repository has received over 400 stars of recognition.
From a technical perspective, the dataset employs images with a resolution of 640x480 and records the seven-dimensional states and action data of the robot's arms and dexterous hands. Unitree has not only open-sourced the complete code for remote operation control but also provided detailed hardware configuration descriptions, installation guides, and data format conversion tutorials.
The G1 robot, which serves as the data source, is a major product launched by Unitree Technology in May of this year, with a starting price of 99,000 yuan. This robot is equipped with a three-finger dexterous hand and utilizes hybrid force-position control technology, making it capable of performing delicate operations such as unscrewing bottle caps, cracking walnuts, and welding. Its integration of Intel RealSense D435 and LIVOX-MID360 3D LiDAR provides it with 360-degree omnidirectional perception capabilities.
By August, the G1 had achieved large-scale mass production. Through deep reinforcement learning and simulation training, the G1's functions have continuously upgraded, allowing it to perform complex actions such as single-leg jumps and 360-degree turns, showcasing strong technical prowess.
The open-sourcing of the G1 dataset by Unitree Technology not only demonstrates its commitment to advancing industry development but also provides valuable resources for technological innovation and ecological construction in the entire robotics industry. With more and more companies joining the open-source movement, the development of humanoid robots is expected to achieve new breakthroughs.
Open-source dataset collection: https://github.com/unitreerobotics/avp_teleoperate
Open-source learning algorithm: https://github.com/unitreerobotics/unitree_IL_lerobot
Open-source datasets and models: https://huggingface.co/UnitreeRobotics
