New technology helps humanoid robots improve 'self-recovery' after falling

Humanup’s randomization improved robustness on varied terrains, achieving 78.3% get-up and 98.3% roll-over success rates. Researchers introduce a breakthrough in robot autonomy A team at the University of Illinois Urbana-Champaign has developed a machine-learning framework that allows humanoid robots to autonomously get up after falling. This is the first successful demonstration of learned fall-recovery strategies for human-sized humanoid robots in real-world conditions. Named HUMANUP, the framework is designed to enhance robot autonomy, making them more adaptable for various applications. “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,” said researchers in the study paper. “This paper develops a learning framework to produce controllers that enable humanoid robots to get up from varying configurations on varying terrains.” Robot fall recovery Humanoid robots frequently fall, limiting their ability to operate independently. The need for human intervention after falls has been a major challenge, especially in complex environments like uneven terrains and confined spaces. The DARPA Robotics Challenge highlighted this issue, as 26 out of 46 trials involved falls, with most requiring human assistance. Researchers explained that designing a fall-recovery controller is challenging due to several factors:

• The irregular movement patterns of falling robots.
• Varied contact points with the ground.
• Sparse reward signals that make it difficult for the robot to learn effective strategies.

Unlike locomotion, where movements follow a predictable cycle, fall recovery requires the robot to determine how to use its limbs and surroundings to stand up. To address this, researchers developed HUMANUP, a learning-based framework that enables humanoid robots to recover from falls in diverse environments. How HUMANUP works The HUMANUP RL framework operates in two stages:

• Discovery phase – The system explores different limb movements to find effective ways to stand up, without prioritizing smoothness or speed.
• Refinement phase – The successful movements are then fine-tuned into controlled and efficient motions, ensuring real-world applicability across different fall scenarios and terrains.

Adaptive humanoid motion To test HUMANUP, researchers used a Unitree G1 humanoid robot on six different types of terrain:

• Concrete
• Brick
• Stone tiles
• Muddy grass
• Grassy slopes
• Snow

Each terrain posed unique challenges, such as varying slopes, stiffness, bumpiness, and roughness. The robot was tested in two key scenarios:

• Rolling over from a prone to a supine position.
• Standing up from a supine position.

HUMANUP’s performance was compared to two other controllers:

• A version of HUMANUP without posture randomization.
• The default controller of the Unitree G1 robot.

Key findings The default G1 controller struggled on rough terrains, slopes, and slippery surfaces like snow. It relied heavily on hand movements, which led to motor overheating and inefficiencies. Standing up took over 11 seconds. In contrast, HUMANUP optimized leg movements, reducing standing time to about six seconds. HUMANUP’s randomization strategy improved its adaptability, achieving:

• 78.3% success rate in standing up.
• 98.3% success rate in rolling over.
• Limitations and future improvements

Despite its success, HUMANUP has some limitations:

• It relies on powerful physics simulations for accurate modeling.
• Ensuring human-like motion remains challenging due to reinforcement learning constraints.
• However, researchers claim that HUMANUP outperforms existing controllers in efficiency and flexibility.

The full research findings are available on the arXiv preprint server.