A research team from the Department of Mechanical Engineering at the Korea Institute of Advanced Technology (KAIST) has developed a technology to control a four-legged robot that can dexterously navigate deformable terrain such as a sandy beach.
The team developed force modeling technology and a neural network capable of making real-time decisions to adapt to different types of ground surface while the robot is moving, and applied it to reinforcement learning.
The trained neural network controller is expected to expand the scope of four-legged walking robots by proving its resilience in changing terrain, including the ability to move at high speed even on a sandy beach, as well as walk and turn on soft ground without losing balance.
This study is published (https://www.science.org/doi/10.1126/scirobotics.ade2256) in Science Robotics under the title "Learning quadrupedal locomotion on deformable terrain".
Reinforcement learning is an AI learning method in which an agent interacts with the environment and uses this set of data to complete a task. Since the volume of data required for reinforcement learning is very large, the method of data collection through virtual simulations is widely used.
In particular, learning-based controllers in the field of walking robots have been applied in a real environment after being trained based on data collected in simulations. But since the performance of a learning-based controller quickly degrades when the real environment has any discrepancies with the learned simulated environment, it is important to implement a real-life-like environment during the data acquisition phase. Therefore, in order to create a learnable controller capable of maintaining balance on a deforming surface, the simulator must provide a similar contact experience.
The research team defined a model that predicts the force generated by ground contact from the motion dynamics of a walking body based on a ground reaction force model that takes into account the additional effect of the mass of the granular medium. In addition, due to the calculation of the force arising from one or more contacts at each step, it was possible to effectively simulate the deformation of the relief.
The research team also implemented an artificial neural network framework that implicitly predicts ground characteristics using a recurrent neural network that analyzes time series data from the robot's sensors.
The created controller was installed on the RaiBo robot, which was built by the research team and demonstrated walking at speeds of up to 3.03 m/s on a sandy beach, where the robot's feet were immersed in the sand. Even when moving on harder terrains such as turf and treadmills, RaiBo was able to run consistently, adapting to the characteristics of the ground without additional programming or overhaul of the control algorithm.
It is expected that the modeling and learning methodology developed by the research team will facilitate the performance of practical tasks by robots.
Share with friends:
Write and read comments can only authorized users