Unitree G1
A mass-produced humanoid robot from Unitree Robotics, featuring high-degree-of-freedom joints and dynamic locomotion capability. It is frequently used in reinforcement learning and robot control research, providing a hardware platform for studying locomotion, whole-body control and manipulation in humanoid embodiments.
Timeline (last 90 days)
A recent study proposes using a diffusion model to directly model the weight distribution of a humanoid robot control policy network, and generated deployable walking, dancing, spinning kick and other action networks on the Unitree G1 robot.
The current work is still limited to the Unitree G1 platform and a specific set of actions, and the generality and scalability of the weight space await more verification.
The research team collected a large number of networks from the reinforcement learning process, and after training the diffusion model, the generated networks can be directly deployed to a real robot to complete diverse motions such as walking, dancing, and spinning kicks, indicating that the generated weights possess actual control capability.
The Pilot paper proposes a unified single-stage reinforcement learning framework for Local Manipulation of humanoid robots in unstructured scenarios, including a cross-modal context encoder and a MoE policy architecture; experiments were conducted in simulation environments and on a 29-degree-of-freedom Unitree G1 physical robot, and the paper claims it outperforms baseline methods in stability, instruction tracking, and terrain traversability.
Unitree G1 is a humanoid robot released by Unitree Robotics, about 1.3 meters tall with more than 23 degrees of freedom, mainly targeting research, education, and lightweight service scenarios.
CSF was demonstrated as a complete system on the real-world Unitree G1 humanoid robot, covering scenarios such as interaction with people and interaction with objects, and successfully prevented unsafe motions.
FAITH's policy was deployed on a real-world Unitree G1 humanoid robot for demonstration.
All Facts (7)
A recent study proposes using a diffusion model to directly model the weight distribution of a humanoid robot control policy network, and generated deployable walking, dancing, spinning kick and other action networks on the Unitree G1 robot.
50%UnverifiedThe current work is still limited to the Unitree G1 platform and a specific set of actions, and the generality and scalability of the weight space await more verification.
50%UnverifiedThe research team collected a large number of networks from the reinforcement learning process, and after training the diffusion model, the generated networks can be directly deployed to a real robot to complete diverse motions such as walking, dancing, and spinning kicks, indicating that the generated weights possess actual control capability.
50%UnverifiedThe Pilot paper proposes a unified single-stage reinforcement learning framework for Local Manipulation of humanoid robots in unstructured scenarios, including a cross-modal context encoder and a MoE policy architecture; experiments were conducted in simulation environments and on a 29-degree-of-freedom Unitree G1 physical robot, and the paper claims it outperforms baseline methods in stability, instruction tracking, and terrain traversability.
50%UnverifiedUnitree G1 is a humanoid robot released by Unitree Robotics, about 1.3 meters tall with more than 23 degrees of freedom, mainly targeting research, education, and lightweight service scenarios.
50%UnverifiedCSF was demonstrated as a complete system on the real-world Unitree G1 humanoid robot, covering scenarios such as interaction with people and interaction with objects, and successfully prevented unsafe motions.
50%UnverifiedFAITH's policy was deployed on a real-world Unitree G1 humanoid robot for demonstration.
50%