AI-Driven Recovery Controller for Quadruped Robot
Author Information
Author(s): Kang Han Sol, Lee Hyun Yong, Park Ji Man, Nam Seong Won, Son Yeong Woo, Yi Bum Su, Oh Jae Young, Song Jun Ha, Choi Soo Yeon, Kim Bo Geun, Kim Hyun Seok, Choi Hyouk Ryeol
Primary Institution: Sungkyunkwan University
Hypothesis
Can a single neural network model enable a quadruped robot to recover from falls in various postures?
Conclusion
The proposed method allows a quadruped robot to successfully recover from falls within 5 seconds, even from completely inverted positions.
Supporting Evidence
- The robot can recover from falls caused by disturbances.
- The method condenses fall recovery knowledge into a single neural network.
- The robot successfully recovers from various postures, including completely inverted ones.
- The approach shows potential for application in other legged robots.
Takeaway
This study shows that a robot can learn to get back up after falling using a smart computer program, even if it falls in tricky ways.
Methodology
The neural network was trained using reinforcement learning in simulations before being applied to a real quadruped robot.
Limitations
The method may not effectively handle all fall scenarios, particularly those involving interference between leg links.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website