It is all Reinforcement learning at Facebook’s New Robotics Lab

Facebook's latest offering, its new nascent robotic platform has been creating quite a stir in the technology circles. Located at its new lab in its palatial Silicon Valley HQ, Facebook developers have been working on the new red and black Sawyer robot arm trying to make it wave all over the place with a mechanical whine. The Sawyer robot is supposed to casually move its hand to a spot in space to its right, to finally reset it to its starting position. This quick movement may not happen that smooth, but the developers are trying their best to achieve the dexterity.

The social media giant is investing heavily into robotics that it thinks holds the key not only for a better robotic future, but also for developing better artificial intelligence. Facebook's latest offering is powered with reinforcement learning, teaching itself to explore the world.

Currently, Facebook is exploring ways that could help a machine learn without much of inputs. These AI-powered machines require a lot of data as inputs to complete a task which may seem simple to a human, like image identification.

The Long-Term Strategy

The team comprising of engineers who are working at Facebook's headquarters at Menlo Park-based Facebook Inc. hope to understand how artificial intelligence might eventually teach itself how to police content on the social network. This is a long-term strategy what the social media giant is expecting to fulfil from its robotic initiatives.

The social media giant has unveiled three of its AI robotics projects which it is currently working on. In one project, researchers have analysed how a six-legged, off-the-shelf robot will teach itself how it can walk. Facebook wants to focus its robotic developments on self-supervised learning or reinforcement learning, a concept in which systems learn directly from raw data to adapt to new tasks and new circumstances through trial and error concept using direct input from sensors.

Reinforcement learning comes at an expense as teaching a machine, especially a robot to learn something on its own is difficult. With the unavailability of training data AI-powered robots, like kids learn through trial and error. As there is unpredictability which exists, machines have to adapt to new situations to adjust to the change.

Self-Supervised Learning

The concept of reinforcement learning or self-supervised learning differs from the traditional way researchers teach AI a task like labelling manual contents, to subsequently find more content that is similar. Though trained AI will learn faster, it is more rigid and has greater probability to fail when it encounters a scenario which it is not being trained for.

For robots, self-supervised learning would mean that robots master the art of self-learning that comes from trying to walk in snow, picking up heavy object in its own without any manual intervention.  Currently, the scope of reinforcement learning is very limited, and it can be said that the robots are very dumb as they need a line of code to be written for each and every action of theirs, whether it is how to walk or moving arm.

LeCun, the chief AI scientist at Facebook says that Facebook is trying to make machines learn which is the next challenge and would make significant progress in AI if achieved successfully. AI research scientist at Facebook, Franziska Meier says that Facebook is experimenting with reinforced learning, and aims to try out instilling a notion of curiosity, among its users.

In a crux, the ultimate idea is to make machines more flexible and less single-minded about a task. AI is undoubtedly making robots smarter, but it is the robots who are also helping in the advancement of AI. There is a lot of the interesting questions and interesting problems that need to be answered which are connected with AI especially to the future state of AI, how can one get to the human-level AI is what is currently being addressed by specialists who are on a mission to bring a new era into reinforcement learning backed robotics.

Who knows, maybe in the future, a robot will know the direction it needs to head to find the exit, in a confusing maze learning through its past experiences leveraging reinforcement learning!