Researchers at the University of Illinois Urbana-Champaign have created a system that helps humanoid robots stand up independently after falling.
Humanoid robots, designed with a human-like structure, are becoming more capable of handling real-world tasks as advancements in control algorithms improve their speed and movement complexity. However, they are prone to tripping or colliding with objects since they walk and run on two legs like humans. Unlike humans, who can easily stand up after falling, these robots often struggle to recover without external assistance.
Researchers at the University of Illinois Urbana-Champaign have developed a machine-learning framework enabling humanoid robots to recover after falling autonomously. The framework could enhance robot autonomy and support their broader deployment.
The research team developed a framework called HUMANUP, which uses reinforcement learning (RL) to help humanoid robots stand up on their own, regardless of their falling position.
Previous applications of humanoid locomotion learning have been successful, but the task of getting up presents additional challenges due to complex contact patterns. Accurately modelling collision geometry and handling sparse rewards are essential for effective recovery. A two-phase approach is used to address these challenges, following a structured curriculum.
The HUMANUP RL framework operates in two stages. The first stage identifies effective limb trajectories that enable a robot to stand up without strict constraints on movement smoothness or speed.
In the second stage, the framework refines the initial motions, transforming them into smooth and controlled movements that the robot can perform. These motions remain effective regardless of the robot’s position or the terrain where it falls.
The researchers tested HUMANUP in simulations and real-world settings using the Unitree G1 humanoid robot, which Unitree Robotics developed. Their results were promising, showing that the robot could autonomously recover from falls, regardless of its position or the surface beneath it.
The framework developed by the team could be refined and adapted for other humanoid robots, enabling them to recover autonomously after falling. This advancement could enhance robot capabilities and support their broader adoption in real-world applications.
