The Everyday Robot Project Aims to Expand Usefulness of Robots
A team of roboticists, computer scientists, researchers, makers and builders at X, Alphabet’s moonshot factory, announced the first results of its “Everyday Robot Project,” which aims to bring robots out of structured environments and into the dynamic, everyday environments of our homes and workplaces.
The team’s project lead, Hans Peter Brondmo, shared details of the team’s progress on a Medium blog post. For the past few years, the project team has been working to determine whether it was possible to create robots that could be helpful, affordable, and able to help people with everyday tasks. The team’s experimental system is exploring three questions:
- Can we create a robotic system that’s able to do useful tasks in the real world that integrates all of the robot’s capabilities – seeing, “understanding,” moving about and picking up or moving objects around?
- Can we build robots capable of quickly learning how to perform new useful tasks through practice instead of having engineers hand code every new task, exception, or improvement?
- Is it possible for a robot to take what is learned from one task and apply that learning to a new task without having to rebuild the robot or create an entirely new application?
Brondmo, in his blog post, said recent advances in machine learning, combined with sophisticated sensor technology and low-cost hardware, “mean that we are much closer than ever to robots becoming mainstream.”
Teach, don’t code
Brondmo said in order for this to happen, there needs to be a shift away from programming a robot, and teaching them instead:
For robots to be useful in everyday environments we need to move away from painstakingly coding them to do specific and structured tasks in exactly the right way at exactly the right time. We have concluded that you have to teach machines to perform helpful tasks; you cannot program them.
The team chose a task (sorting waste from recyclables) that was complex enough to where the team wasn’t sure whether it could be done, but not so difficult that it would take a year in order to determine whether it was possible.
Simulation to reality
In order to achieve this, the team used several machine learning techniques, including simulation, reinforcement learning, and collaborative learning, Brondmo reported. “Each night, tens of thousands of virtual robots practice sorting the waste in a virtual office in our cloud simulator; we then move the training to real robots to refine their sorting ability,” Brondmo wrote. “This real-world training is then integrated back into the simulated training data and shared back with the rest of the robots so that the experience and learning of each robot is shared with them all.”
Over the last few months, the robots sorted thousands of pieces of waste, reducing the office’s waste contamination levels from 20% to less than 5%. Brondmo said the project showed they could create a robotic system that integrates the robot’s capabilities to do something generally useful, as well as prove that it’s possible for robots to learn how to perform new teasks in the real world through practice, rather than the programming for every new task.
Brondmo said the next challenge for the team is to see whether they can take what the robot learned from this task, and apply it to another task without rebuilding the robot or “writing a ton of code from scratch.” He admits that “this could prove to be impossible, but we’ll give it a shot.”
The project team is promising more updates on its project page.