A team of researchers led by Professor Jemin Hwangbo from the Department of Mechanical Engineering successfully developed a four-legged robot that could robustly navigate unsteady and deformable walking environments, such as sandy beach coasts or fluffy bed mattresses. The research was published last January in Science Robotics, and received support from the Samsung Research Funding & Incubation Center of Samsung Electronics.

Multipedal robots have already been performing superbly on various walking surfaces through the use of artificial intelligence (AI). However, since robot AI models tend to underperform on surfaces that they have not learned before, it is important that their training environments simulate several arbitrary situations that they are likely to encounter. As such, the team used reinforcement learning to collect data through approximated simulations of real physical phenomena and train their robot model to perform certain tasks. They further expanded from existing ground reaction models — models that calculate the amount of force produced upon contact of the foot on the ground during walking — to take into consideration deformable terrains, which helped the team simulate such environments for the robot. The team then used an artificial neural network to allow real-time decision-making for the robot. Specifically, the neural network takes in time-series data from the robot’s built-in sensors, which are used to predict characteristics of the current terrain so that the robot can adjust accordingly.

The robot, named “RaiBo”, could walk as fast as three meters per second on sand even while its feet were completely buried in the sand, and could rotate at almost 90 degrees per second on an air mattress without falling over. Furthermore, the robot showed impressive stability even on harder common terrains like a running track, proving that it is capable of adjusting to multiple environments without the need for manual or direct readjustments to the algorithm. With this developed method of integrating the capability to readjust to different terrains, the team hopes to expand the scope of practical tasks that automated robots could perform.

In an interview with the research paper’s first author, Suyoung Choi, he reiterated the main points of their study and its benefits, stating that “providing a learning-based controller with a close contact experience with real deforming ground is essential for application to deforming terrain”. He further adds that “the proposed controller can be used without prior information on the terrain, so it can be applied to various robot walking studies.”