Trip-proof robot adapts to challenging terrain in real time
Improvising is hard for robots, and hitting an unusual surface or obstacle usually means an abrupt stop or a hard fall. But researchers have created a new locomotion model that adapts in real time to any terrain the robot encounters, changing its gait on the go so it can keep moving when it meets sand, rocks, stairs and other sudden changes.
Although robot movement can be versatile and precise, and it can learn to climb steps and cross rough ground, such behavior is more like individually trained skills that the robot switches between. Well-known robots such as Spot can recover from pushes or kicks.
In fact, the system simply works to correct a physical anomaly while sticking to an unchanged walking policy. There are several adaptive motion models, but some are very specific, and others need enough time to work, so the robot is bound to fall by the time they take effect.
A team from Facebook AI, the University of California, Berkeley and Carnegie Mellon University named the system Rapid Motor Adaptation. It was created because people and animals can quickly, efficiently and unconsciously change their gait depending on circumstances.
“Suppose you are learning to walk and go to the beach for the first time. Your foot sinks, and to pull it out you need to apply more force. It feels strange, but after a few steps you walk naturally, as you would on solid ground. What is the secret?”
Of course, if you have never encountered a beach before, at that moment you do not switch into some special "sand mode" that lets you walk on soft surfaces. The way you change your movement happens automatically and without any real understanding of the external environment.
The human body responds to different physical conditions, sensing the different effects those conditions have on the body itself. The RMA system works in a similar way: while moving in new conditions, the walking rules change within a very short time (half a second or less).
The system was fully trained by simulation in a virtual version of the real world, where the robot's small brain (everything runs locally on an onboard limited computing unit) learned to maximize forward motion with minimal energy and avoid falls, immediately observing and responding to data coming from its accelerometers and other physical sensors.
To highlight the fundamentally internal nature of the RMA approach, Malik said the robot does not use visual input at all. If people and animals without sight can walk normally, why shouldn't a robot be able to? It is impossible to estimate "external effects" such as the exact friction coefficient of the sand or rocks it walks on, so it simply pays close attention to itself.
Ultimately, the system consists of two parts: a main, always-running algorithm that actually controls the robot's gait, and an adaptive algorithm running in parallel that monitors changes in the robot's internal readings. When significant changes are detected, the latter analyzes them, roughly "the legs should do this, but they are doing that, so the situation must be such-and-such", and tells the main model how to adjust.
From this point on, the robot only thinks about how to move forward in these new conditions, effectively improvising a special gait. After simulation training, the robot had already succeeded in the real world.
The press release says: "The robot was able to walk on sand, mud, hiking trails and tall grass without a single failure in all of our trials. The robot successfully descended stairs on a hiking trail in 70% of trials. It successfully crossed a pile of cement and a pile of pebbles in 80% of trials, despite never having seen unstable or sinking ground, obstructing vegetation or stairs during training. It also maintained its height with a high degree of success while carrying a 12 kg payload, which was 100% of its weight."
Malik endorsed research by New York University professor Karen Adolph, whose work showed how adaptable and unscripted the process of learning to walk is for humans. The team's goal is that if you need a robot that can handle any situation, it should learn to adapt from scratch rather than having many modes to choose from.
Just as it is impossible to build a smarter computer vision system by exhaustively labeling and documenting every object and interaction, you cannot prepare a robot for a diverse and complex physical world with 10, 100 or even thousands of special walking parameters for gravel, mud, rubble and wet wood.
The core of the system could potentially be applied not only to legged robots but also to entirely different areas of AI and robotics.
In addition, combining base and adaptive algorithms could work for other intelligent systems. Smart homes and municipal systems typically rely on existing policies, but what if they could adapt on the go?
