Alex Wilkins is a writer.
A robot dog can learn to walk on difficult terrain in just 20 minutes thanks to a machine learning algorithm.
When they encounter unfamiliar environments, they tend to struggle, because they have to be carefully programmed by humans or extensively tested in simulations before they can perform real-world tasks.
A robot using a kind of machine learning called deep reinforcement learning can work out how to walk in about 20 minutes in various environments, such as a grass lawn, a layer of bark, and a memory foam.
A model of the target terrain is not required for the robot to use Q- learning. Machine learning is used in simulations Levine says that we don't need to understand the physics of an environment to turn a robot on.
The robot gets a reward for every action it performs, depending on how successful it was. This process is repeated continuously until it learns to walk.
The team member says it is similar to how people learn. Think about your previous experience and try to understand what could have been improved by interacting with an environment.
Levine says the team would need to fine- tune the model's reward system if the robot is to learn other skills.
The amount of variables and data that have to interact at the same time makes it difficult to make deep reinforcement learning work.
Watkins thinks it's very impressive. I was surprised that Q-learning could be used to learn skills such as walking on different surfaces with little experience and in real time.
There is a reference on arxiv.org.
There are more on this topic.