Researchers at Google’s AI lab have developed an AI system that learns from the motions of animals. The team has developed a dog-like robot quadruped that learns dog behavior by studying how real dogs move.
This study is interesting is because teaching a robot by showing it videos of a real dog is a very difficult task, so far now we were able to train the robot to perform tasks by mimicking the movements of a living creature, for eg. robot arms that build cars are taught how to spot weld or tighten bolts by mimicking the desired actions.
The aim of this research was to figure out how to efficiently and automatically transfer “agile behaviors” like a light-footed trot or turn from their source (a great dog) to a quadrupedal robot.
By using the dataset of motion recorded from various sensors attached to a dog, the researchers taught Laikago the quadruped robot many different movements that are hard to learn through traditional methods.
The motion data from real dog then get used to construct simulations of each maneuver, including a dog trot, side-step, and … a dog version of classic ’80s dance move, the running man.
The team of researchers then matched together key joints on the simulated dog and the robot to make the simulated robot move in the exact same way as the animal.
The teams’ framework takes a motion capture clip of an animal like a dog, in this case, and uses reinforcement learning, a technique that spurs software agents to complete goals via rewards, to train a control policy.
Providing the system with different reference motions enabled the researchers to teach a four-legged Unitree Laikago robot to perform a range of behaviors, from fast walking to hops and turns.
To work out possible interpretation errors (because the digital dog is made from metal and motors instead of bones, muscles, and sinews), the team shows the AI system multiple stop-action videos of a real dog in action.
The AI system builds up a toolset of possible moves depending on scenarios that might be encountered in the real world. Once the simulation has built up a knowledge base, its “brain” is uploaded to Laikago, who then uses what the simulation has learned as a starting point for its own behavior.
But simulators generally provide only a coarse approximation of the real world. To address this issue, the researchers employed an adaptation technique that randomized the dynamics in the simulation by, for example, varying physical quantities like the robot’s mass and friction.
These values were mapped using an encoder to a numerical representation i.e., an encoding that was passed as an input to the robot control policy. When deploying the policy to a real robot, the researchers removed the encoder and searched directly for a set of variables that allowed the robot to successfully execute skills.
The team says they were able to adapt the policy to the real world using under eight minutes of real-world data across approximately 50 trials.
Moreover, they demonstrated that the real-world robot learned to imitate various motions from a dog, including pacing and trotting, as well as artist-animated keyframe motions like a dynamic hop-turn.
“We show that by leveraging reference motion data, a single learning-based approach is able to automatically synthesize controllers for a diverse repertoire [of] behaviors for legged robots,” wrote the coauthors in the paper. “By incorporating sample efficient domain adaptation techniques into the training process, our system is able to learn adaptive policies in a simulation that can then be quickly adapted for real-world deployment.”
Link to the paper:
More in AI