Scientists have unveiled a robotic dog that can teach itself to walk in just an hour.
In a video released by researchers, the four-legged robot can first be seen kicking its legs in the air and struggling – but after just 10 minutes it can take steps – and after an hour it’s walking fairly easily and rolling off his back and even being knocked over by one of the explorers with a stick while navigating.
Unlike many robots, this one has not been shown beforehand what it has to do in a computer simulation.
Danijar Hafner, an artificial intelligence researcher at the University of CaliforniaBerkeley, worked with his colleagues to train the robot using reinforcement learning.
Scroll down for videos
A robotic dog has been trained to run, roll over obstacles and navigate in about an hour, researchers at the University of California, Berkeley reveal. Pictured above, the robot at the five minute mark
“Typically, robots learn through a large amount of trial and error within computer simulations that are much faster than real-time,” Hafner told DailyMail.com via email.
“After a task such as getting up and walking has been solved in the simulation, the learned behavior is then carried out on a physical robot.
“But simulations can’t capture the complexity of the real world, so behavior that works well in simulations may not solve the task in the real world.”
This type of machine learning is all about training algorithms by rewarding them for performing specific actions in their environment.
Hafner and his collaborators – Philipp Wu and Alejandro Escontrela – used an algorithm called Dreamer, which creates a model of the real world based on past experience and also allows the robot to perform trial-and-error calculations.
The researchers used an algorithm called Dreamer, which uses past experiences to create a model of the real world for the robot to learn from. Pictured above is the robot at 30 minutes
“The Dreamer algorithm has recently shown great promise for learning from small sets of interactions by planning within a learned world model,” the researchers state in their article paperwhich has not yet been reviewed by experts.
“Learning a world model to predict the outcomes of potential actions allows for imaginative planning and reduces the amount of trial and error required in the real-world environment.”
‘Reinforcement learning will be a cornerstone in the future of robotic control,’ said a non-study scientist. Pictured above is the robot at 40 minutes
After an hour, the robotic dog pictured above is quite good at navigating its surroundings, turning around, and more
After the robot learns to walk, it could also learn to adapt to other, less predictable outcomes — such as being poked with a stick by researchers.
Even with reinforcement learning, which has brilliantly outperformed humans in things like board or video games, the world of teaching robots to behave properly in the real world is extremely challenging – as engineers have to program whether or not every action is rewarded, depending on whether it is desired by scientists.
“Applying reinforcement learning to physical robots is challenging because we can’t speed up time in the real world, and robot simulators often don’t capture the real world accurately enough,” Hafner and his colleagues told DailyMail.com.
“While Dreamer shows promising results, learning on hardware over many hours causes wear and tear on robots that may require human intervention or repair,” the researchers note in the study. Pictured above, the robot navigates over an obstacle
“Our project has shown that learning world models can drastically accelerate the learning of robots in the physical world.
“This brings reinforcement learning closer to solving complex automation tasks, such as manufacturing and assembly tasks and even self-driving cars.”
“A roboticist has to do this for every single task [or] problem that the robot is designed to solve,” explains Lerrel Pinto, an assistant professor of computer science at New York University who specializes in robotics and machine learning MIT Technology Review.
That would result in an extensive amount of code and a number of situations that just aren’t predictable.
The research team cites other barriers to this type of technology:
“While Dreamer shows promising results, learning on hardware over many hours causes wear and tear on robots that may require human intervention or repair,” the study’s abstract reads.
“Also, more work is needed to push Dreamer’s limits and our baselines through longer training.
‘Finally, we see tackling more challenging tasks, possibly by combining the advantages of rapid real-world learning with those of simulators, as a powerful future research direction.’
Hafner hopes to teach the robot how to obey spoken commands and perhaps connect cameras to the dog to give it a view – all of which would allow it to perform more typical canine activities like fetching.
In a separate study, researchers at Germany’s Max Planck Institute for Intelligent Systems (MPI-IS) have shown in new research that their robotic dog, named Morti, can easily learn to walk using a complex algorithm that includes sensors in its feet.
“As engineers and roboticists, we sought the answer by building a robot that has reflexes like an animal and learns from mistakes,” says Felix Ruppert, former PhD student in the Dynamic Locomotion research group at the MPI-IS, in a expression.
“If an animal stumbles, is that a mistake? Not once it happens. But if it stumbles a lot, that gives us a measure of how well the robot is running.”
The robot dog works with a complex algorithm that controls how it learns.
Information from foot sensors is matched with data from the machine’s spinal cord model, which runs as a program in the robot’s computer.
The robot dog learns to walk by constantly comparing set and expected sensor information, going through reflex loops and adapting its movement regulation.
Scientists at the Max Planck Institute for Intelligent Systems in Germany have used algorithms to teach a robotic dog named Morti to walk
