Robo-Dog Seeks Partner for Long Walks on the Beach

Quadrupedal robots are becoming an ever more common sight in a variety of industries because of their unique capabilities. These robots are designed to mimic the movement and stability of animals, making them well-suited for tasks such as search and rescue, industrial inspection, and more.

One of the primary advantages of quadrupedal robots is their ability to navigate difficult terrain. With four legs, these robots can easily navigate obstacles and uneven surfaces, making them ideal for search and rescue operations in rugged terrain. They can also be used in industrial settings, such as in inspecting pipelines and other infrastructure in hard-to-reach areas.

However, the real world has a seemingly endless variety of terrain types, and designing a quadrupedal robot that is sufficiently versatile to traverse them all is exceedingly challenging. Moreover, certain environments, like a sandy beach, are difficult to master even if the robot is built with them in mind.

An engineering lab at the Korea Advanced Institute of Science & Technology has put forward a new approach that looks like it may help four-legged robots find their footing, even when the going gets tough. They have developed a novel reinforcement learning-based method to help robots learn to navigate previously unseen terrain, and they have paired this method with an adaptive control architecture that allows the robot to identify the properties of the terrain through touch. Their techniques were implemented in the dog-like Raibo robot that was developed in-house.

Reinforcement learning has previously produced some excellent results by teaching machines to perform a desired task by rewarding good outcomes, and punishing negative results. But, this type of learning requires a very large amount of data to train an accurate model. As such, it is very common to use simulated environments to collect the data and make this technique practical.

However, when the task is complex, like learning to walk on an uneven, sandy surface, if the simulated environment differs from reality in even small ways, it can produce a disastrous result when the model is unleashed in the real world. To deal with this challenge, the team designed their own ground reaction force model that predicts the forces generated upon contact of the walking robot with the surface it is treading on. Using this model to determine contact forces, they were able to very accurately simulate even complex, deforming terrain like sand.

The more accurate simulation environment produces a robot uniquely equipped for new environments, but the team also developed a recurrent neural network that analyzes time-series data from the robot's sensors to predict ground characteristics in real time. This further improves the agility of the system and helps it adapt to each situation it encounters.

Using this innovative technique, Raibo was able to run along the beach at a rate of nearly ten feet per second. Impressively, this was in an environment where the robot’s feet were almost entirely submerged beneath the sand. And this dog was no one-trick pony — it could also run on harder surfaces, like grassy fields and a running track, without any additional training or other modifications.

The team hopes that their system will be used in the future to produce quadrupedal robots that are more agile, and that can be used in a much wider range of applications. That sounds like a realistic goal, and building those robots should also prove to be simpler using the methods that they have described.