Humanoid robots can swiftly get up after they fall with new learning framework

Humanoid robots, which have a body structure that mirrors that of humans, could rapidly and effectively tackle a wide range of tasks in real-world settings. These robots and their underlying control algorithms have improved considerably in recent years. Many of them can now move faster, emulating various human-like movements.

As these robots are designed to walk or run similarly to humans, thus balancing on two legs, they can sometimes collide with objects or trip on uneven terrain, falling to the ground. Yet, in contrast with humans, who can easily pick themselves up when they fall, humanoid robots can sometimes get stuck on the ground, requiring the support of human agents to get back on their feet.

Researchers at the University of Illinois Urbana-Champaign recently developed a new machine learning framework that could allow humanoid robots to automatically get back up and recover after falling to the ground. This framework, presented in a paper on the arXiv preprint server, could make these robots more autonomous, potentially contributing to their future large-scale deployment.

"Hand-designing controllers for getting up is difficult because of the varied configurations a humanoid can end up in after a fall and the challenging terrains humanoid robots are expected to operate on," wrote Xialin He, Runpei Dong and their colleagues in their paper. "This paper develops a learning framework to produce controllers that enable humanoid robots to get up from varying configurations on varying terrains."

The new framework developed by this research team, dubbed HUMANUP, relies on a reinforcement learning (RL) approach. This approach is designed to improve the ability of humanoid robots to get up, irrespective of their position when they fall.

"Unlike previous successful applications of humanoid locomotion learning, the getting-up task involves complex contact patterns, which necessitates accurately modeling the collision geometry and sparser rewards," wrote He, Dong and their colleagues. "We address these challenges through a two-phase approach that follows a curriculum."

The HUMANUP RL framework spans across two different stages. During the first stage, the framework focuses on identifying good limb trajectories that would allow a robot to get up, which pose minimal constraints on how smooth the robot's movements should be or the speed with which these movements should be executed.

During the second phase, on the other hand, the framework refines the motions uncovered as part of the earlier phase, ultimately turning them into smooth and slow motions that can be performed by the robots. Notably, these refined motions should also be effective irrespective of the position of the robot and the terrain on which it fell.

The researchers tested their framework in both simulations and real-world environments, deploying it on the Unitree G1 humanoid robot, an advanced robotic system created by the Chinese company Unitree Robotics. Their findings were highly promising, as they found that their approach allowed the robot to autonomously recover after falling, irrespective of the position it was in and the terrain beneath it.

"We find these innovations enable a real-world G1 humanoid robot to get up from two main situations that we considered: a) lying face up and b) lying face down, both tested on flat, deformable, slippery surfaces and slopes (e.g., sloppy grass and snowfield)," wrote He, Dong and their colleagues. "To the best of our knowledge, this is the first successful demonstration of learned getting-up policies for human-sized humanoid robots in the real world."

The new promising framework developed by He, Dong and their colleagues could soon be further improved and deployed on other humanoid robots, equipping them with the ability to automatically get themselves back up after falling. This could help to further advance the robots, which could facilitate their future widespread adoption.

© 2025 Science X Network