How an AI placed between a front camera and four legs is enough to make an all-terrain autonomous robot

He climbs a staircase with irregular steps, crosses an embankment while the soft ground slips away under his feet, progresses on loose planks without losing his balance and without hesitation… This little quadruped robot reacts quickly to the world around him. and performs movements of surprising agility, almost animal.

Here is the result of three years of research on vision-assisted locomotion, carried out by a joint team from Carnegie Mellon University and the University of California at Berkeley. The results are the subject of a scientific preprint and will be presented at the International Conference on Robotic Learning (CoRL), which will start in Auckland on December 14, 2022.

A beautiful applied work, which shows that the practice is at least as important as the theorycomments Fabien Moutarde, director of the robotics center at Mines Paris. This is an important advance in the methods used to make a legged robot move in difficult environments, even if there is no particular algorithmic revolution.. »

These researchers are opting for an approach that has been explored for ten years in robotics, based on automatic learning. “ Thanks to intelligent visual controls, the robot is taught to adopt the best posture or gait according to the image it perceives.explains Fabien Mustard. Like humans, who see and deduce by reflex what they must do, whether they are walking or driving a car. »

This is an “end to end” control principle. The usual intermediate steps which are used to calculate displacements – multiple perception, data fusion, representation of the environment, trajectory calculation, etc. – are eliminated here. ” The construction of a map of the world is a complex operation, which requires for example a good odometryemphasizes Fabien Mustard. And these cards are very noisy. »

Reinforcement learning then supervised

Based on the standard A1 platform from the manufacturer Unitree, this adaptive robotic system is based on a recurrent neural network. Its training includes two phases carried out in simulation: learning by reinforcement with a 3D mini-matrix and environmental parameters (friction, etc.), then supervised learning from depth data (acquired by the 3D camera) and proprioception (perception of the position of the parts of the robot).

Once trained and placed in real conditions, the small quadruped automatically activates its joints with the most favorable angle for the terrain crossed, according to depth and proprioception information. As the videos show, it does well in very rough or slippery terrain.

The recurrent neural network has the advantage of storing information from the immediate past. ” The robot reasons – if we can put it that way – in a somewhat temporal way: it remembers what it saw a few moments before “Summarizes Fabien Mustard. Like a cat, which puts its back paw on the exact spot where it has just put its front paw.

The researchers do not hide the possible accidents along the way, such as falling from a step too high for such a short-legged robot (about thirty centimeters at the withers). Moreover, the simulations used for learning cannot reflect the extent of the field of possibilities, infinite in essence. ” If this robot encounters a situation that is too different from what it has learned, perhaps it will behave less well. But it’s true for everything “, acknowledges Fabien Mustard.

Selected for you

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *