A New Learning Framework Enables Humanoid Robots to Quickly Recover and Stand Up After Falling

Humanoid robots, designed to resemble the human body, are becoming increasingly capable of efficiently handling diverse tasks in real-world environments. Advances in their control algorithms have led to significant improvements, enabling many to move faster and mimic human-like motions more effectively.

Since humanoid robots are designed to walk and run like humans, balancing on two legs, they may occasionally collide with objects or trip on uneven surfaces, causing them to fall. Unlike humans, who can quickly get back on their feet, these robots sometimes struggle to recover and require human assistance.

A Machine Learning Framework for Autonomous Recovery

To address this challenge, researchers at the University of Illinois Urbana-Champaign have developed a machine learning framework that enables humanoid robots to autonomously rise and recover after falling. Detailed in a paper on the arXiv preprint server, this framework could enhance robot autonomy and support their large-scale deployment in the future.

“Designing controllers for recovery is challenging due to the numerous positions a humanoid robot can assume after a fall and the complex terrains they must navigate,” wrote Xialin He, Runpei Dong, and their colleagues in their paper. “This study introduces a learning framework that generates controllers, allowing humanoid robots to stand up from diverse positions across different terrains.”

The research team developed a framework called HUMANUP, which utilizes a reinforcement learning (RL) approach. This method enhances humanoid robots’ ability to stand up, regardless of their position after a fall.

“Unlike prior successes in humanoid locomotion learning, the task of getting up involves intricate contact patterns, requiring precise collision modeling and sparser rewards,” wrote He, Dong, and their colleagues. “We tackle these challenges using a two-phase, curriculum-based approach.”

Two-Stage Learning for Effective Recovery

The HUMANUP RL framework operates in two stages. In the first stage, it prioritizes discovering effective limb trajectories that enable a robot to stand up while imposing minimal constraints on movement smoothness or execution speed.

In the second phase, the framework refines the motions identified in the initial stage, transforming them into smooth and controlled movements that the robot can execute. These optimized motions must remain effective regardless of the robot’s position or the terrain where it has fallen.

Real-World Testing on the Unitree G1 Robot

The researchers evaluated their framework in both simulations and real-world scenarios, implementing it on the Unitree G1 humanoid robot, an advanced system developed by the Chinese company Unitree Robotics. Their results were highly promising, demonstrating that the approach enabled the robot to autonomously recover from falls, regardless of its position or the surface it landed on.

“We found that these innovations enable a real-world G1 humanoid robot to stand up from two key positions: (a) lying face up and (b) lying face down. Both scenarios were tested on various surfaces, including flat, deformable, slippery terrains and slopes such as grassy inclines and snowfields,” wrote He, Dong, and their colleagues. “To our knowledge, this is the first successful real-world demonstration of learned getting-up policies for human-sized humanoid robots.”

The promising framework developed by He, Dong, and their team could soon be refined and applied to other humanoid robots, allowing them to autonomously recover from falls. These advancements may accelerate the development of humanoid robots, supporting their broader adoption in real-world applications.

Read the original article on: TechXplore

Read more: Watch: The First Humanoid Robot to Do a Front Flip