New framework enables animal-like agile movements in four-legged robots
What is the purpose of Science X's editorial process?
The purpose of Science X's editorial process is to ensure the credibility, accuracy, and quality of the content published on their platform. This is achieved through rigorous fact-checking, adherence to editorial standards, and multiple rounds of editorial scrutiny by their team of writers and editors3. The process aims to provide readers with reliable, up-to-date, and comprehensive coverage of science, technology, and medical news4.
How does the new framework improve robot agility?
The new framework improves robot agility by constructing knowledge at three levels: primitive, environmental, and strategic. It uses deep generative models to produce motor control signals, enabling legged robots to act like real animals. The framework aligns traversing capabilities with the environment and focuses on complex downstream tasks by reusing knowledge from previous levels. This results in lifelike agility and strategy in robots, allowing them to traverse complex obstacles and adapt to changing environments.
What methods were used to train the quadrupedal robots?
Quadrupedal robots were trained using a new hierarchical framework introduced by researchers at Tencent Robotics X. This framework includes three stages of reinforcement learning, focusing on extracting knowledge at different levels of locomotion tasks and robot perception2. The controller at each learning stage is referred to as PMC, EPMC, and SEPMC. The framework was evaluated on a quadrupedal robot called MAX, enabling it to traverse complex obstacles and mimic animal-like agility.