Self-Teaching Robot Dog Can Walk, Roll Over, Navigate Obstacles in Just One Hour of Training

A robot drog developed by engineers from UC Berkeley has been trained to walk, navigate obstacles, and roll over in the first hour. Using an algorithm, researchers allowed the dog to experience commands before being activated with reinforcement learning as the cornerstone tool for future robot control.

Self-Teaching Robot Dog Learns Tricks in the First Hour

Scientists just unveiled a robotic dog that teaches itself to walk in the first hour. The video shows the four-legged robot flailing its legs at first and struggling. However, after 10 minutes, it takes its first steps. By the first hour mark, it walks easily, rolling on its back and navigating complex obstacles.

Danijar Hafner, a co-author of the study and a researcher of artificial intelligence at UC Berkeley, worked alongside his colleagues to train the robot dog using reinforcement learning.

Robots typically learn tasks and movements via a lot of trial and error in computer simulations that are sped up in real-time. After the robot solves a task, like standing or walking, in the simulation, it is run on the physical robot in the simulation, reports DailyMail.

Hafner explains that simulations cannot capture the complexities of the real world. Hence the behavior that works incredibly in simulation may fail to solve tasks in the real world.

In the initial paper published in Arxiv, titled "DayDreamer: World Models of Physical Robot Learning," Hafner and his colleagues used the Dreamer algorithm that works from past experiences to build a model of the real world while allowing the robot to conduct trial-and-error calculations via a computer program to predict potential future outcomes of its won actions. This allows the robot to learn faster. Once the robot learns to walk, it keeps learning to adapt to various situations.

Robot Learning: Reinforcement Training

Reinforcement learning has been improving robots to outdo humans in board and video games while teaching it how to act properly in extremely challenging real-world scenarios. This requires engineers to program whether an action earned a reward or not based on the desired outcome of the team.

Applying reinforcement learning to real-life robots has been a big challenge for experts. Hafner explains that the team's recent endeavor demonstrates that learning world models drastically hastens the robot's learning in the physical world. This brings evidence of reinforcement learning as a close solution to complex automation tasks like manufacturing, assembly tasks, and even self-driving cars.

Lerrel Pinto, assistant professor of computer sciences at New York University, explains to MIT Technology Review that roboticists must do this for each task or problem they war robots to solve.

Meanwhile, the team behind the robot dog also cites different obstacles to the technology. While the Dreamer algorithm shows promise, learning on hardware over numerous hours creates wear on the robots that may soon require human intervention or repair.

Also, more research and tests are needed to explore the algorithm's limitations and the baseline for a longer training time.

Check out more news and information on Tech & Innovation in Science Times.