KAIST (Chairman Kwang Hyung Lee) announced on the 25th that a research team led by Professor Jemin Hwangbo of the Department of Mechanical Engineering has developed a four-legged robot control technology that can walk powerfully with agility even in deformed terrain such as a sandy beach.
Professor Huangpu’s research team developed a technique to model the force received by a ground-walking robot made of granular materials such as sand and simulate it via a four-legged robot. Also, the team worked on an artificial neural network architecture suitable for making the real-time decisions needed to adapt to different types of land without prior information while walking at the same time and applied it to reinforcement learning. The trained neural network controller is expected to expand the application range of quad-walking robots by proving its robustness in changing terrain, such as the ability to move at high speed even on a sandy beach and to walk and turn on floors as soft as air. tidy without losing balance.
This research with a Ph.D. Student Soo-Young Choi from the KAIST Department of Mechanical Engineering as the first author, was published in January in Robotics science. (Paper title: Learning quadrilateral movement on deformed ground).
Reinforcement learning is an artificial intelligence learning method that is used to create a machine that collects data on the results of various actions in an arbitrary situation and uses this set of data to perform a task. Because the amount of data required for reinforcement learning is so huge, a method of data collection through simulations that approximates physical phenomena in the real environment is widely used.
In particular, learning-based controls in mobile robotics have been applied to real environments after learning through data collected in simulations to successfully perform walking controls in various terrains.
However, since the performance of a learning-based controller rapidly decreases when the actual environment has any discrepancy from the acquired simulated environment, it is important to implement an environment similar to the real environment at the data collection stage. Therefore, in order to create a learning-based controller that can maintain balance in deformed terrain, the simulator must provide a similar contact experience.
The research team identified a contact model that predicted the force generated on contact from the motion dynamics of a walking body based on a ground reaction force model that took into account the additive collective effect of granular media identified in previous studies.
Moreover, by calculating the force generated by one or several contacts at each time step, the deformed topography was efficiently simulated.
The research team also presented an artificial neural network architecture that implicitly predicts land properties using a recurrent neural network that analyzes time series data from the robot’s sensors.
The acquired control was installed on the “RaiBo” robot, which was practically constructed by the research team to demonstrate walking at a high speed of 3.03m/s on a sandy beach as the robot’s feet were completely immersed in the sand. Even when applied to difficult terrain, such as grass fields and running tracks, it was able to run steadily by adapting to the characteristics of the ground without any additional programming or control algorithm revision.
In addition, it rotated steadily at 1.54 radians/sec (about 90 degrees per second) on an air mattress and showed its quick ability to adapt even in a situation where the ground suddenly became soft.
The research team showed the importance of providing an appropriate communication experience during the learning process by comparing it with a controller that assumed the ground to be solid, and proved that the proposed recurrent neural network modifies the way the controller walks according to the characteristics of the ground.
The simulation and learning methodology developed by the research team is expected to contribute to robots performing practical tasks as it expands the range of terrain that different walking robots can operate on.
First author, Suyoung Choi, said, “Providing a learning-based controller with close contact experience with a real deformed ground has been shown to be essential for application to terrain deformation.” He went on to add that “the proposed controller can be used without prior terrain information, so it can be applied to various robotic walking studies.”
This research was conducted with the support of Samsung Research Funding & Incubation Center of Samsung Electronics.