We are moving away from a world where robots have to be painstakingly coded towards a future where robots can learn.
They do what they’ve been programmed to do in exactly the right way, at exactly the right time.
They are useful, but only in very certain circumstances, and are incapable of adapting to the unpredictable and unstructured nature of everyday life.
Today
Operate in environments designed and structured for them, like factories.
Are configured and scripted to perform specific, repetitive and often highly precise tasks. Like assembling a smartphone or placing chips on a computer board.
Extend our capabilities and increase our productivity by helping us with hard jobs. Such as lifting a big, heavy car door into place on an assembly line.
But what they can’t do is adapt to the unpredictability of real life. They can't help us with everyday tasks, or navigate unstructured environments.
Tomorrow
Assist us in everyday environments where things are constantly changing. They’ll support us in the places we work, in our hospitals, in our communities — and eventually in our homes.
Learn how to assist us with an ever-growing range of tasks, in known and unknown environments.
Understand our preferences — our likes, our needs — and how they can lend us a helping hand when we need it most.
Supporting us in the unstructured, unpredictable spaces where we spend our time, and putting the benefits of robotics into the hands of everyday people.
Put simply, we’re moving from a world where code and programming determine how robots sense, plan and act to one where data and learning do.
Learning is our moonshot. We believe it's the key to unlocking generalization, enabling robots to perform an infinite number of tasks in near infinite environments, and solving the small problems that take up much of our daily lives in the process.
To make our vision a reality, we have developed an integrated hardware and software platform that makes learning feasible, at scale.
Let’s start by meeting the body of a helper robot — the first thing you see — the ‘Hardware’.
Robots don’t need to look like us to help us. That’s why our robots have wheels, not legs. In an increasingly accessible world, wheels allow robots to go almost anywhere a person can.
Our end of arm system is designed to take on multiple tasks, like picking up objects, wiping surfaces, and opening doors or drawers. To allow for this flexibility, our robots’ arms aren't like a person’s, they’re modular. Tools can be added or removed depending on what needs to be done: grippers for grasping, squeegees for wiping, brushes for dusting, and more.
To understand the world, our robots collect real-time visual and spatial data through a sophisticated suite of integrated cameras, lidar, IMUs and bumper sensors. Together, this suite allows them to process color, depth, time-of-flight, and other information to create a detailed map of their environment as they explore it.
Software
Let’s move onto the brain of a helper robot — the ‘Software’. Hardware allows our robots to perceive and move in real-world spaces, while machine-learning enables them to adapt to the unpredictability of real life.
Our robots are learning machines
Our robots explore the world and, through experience and reinforcement, get better at doing the things that everyday life requires.
To learn in the real-world, a robot needs to be able to sense and understand it. Hardware collects visual data, and builds a detailed map of space. Software allows the robot to translate what it’s seen (a sea of 1’s and 0s) into an understanding of an environment through object and people segmentation. Through supervised learning, the robot can tell that thing over there is a desk, or a can. The more it sees, the better its perception system gets at segmentation and classification, enabling it to tell an ever-growing amount of objects apart.
Taking on new tasks
After a robot understands its surroundings, it can plan and act in the world using its motion system. But, just like us, robots don’t know how to do a new task immediately. This is where imitation and reinforcement learning come in.
Imitation learning teaches a robot tasks through modeled behavior. Using machine learning and human demonstration, a robot is shown how to perform a task. The robot then mimics this behavior, jumpstarting its learning in unfamiliar environments.
Meanwhile, reinforcement learning rewards a robot everytime it successfully completes a task, and informs it when it fails. Over time, the actions that lead to success get rewarded more often, and the robot becomes more skilled. This information is then shared with other robots through our integrated platform, allowing the fleet to improve as a whole.
No robot is totally self-reliant, so to accelerate their learning, we lend them a helping hand. Our team of robot operators provide guidance through demonstration, interaction and real time feedback so our robots can learn to adapt to a breadth of conditions.
Like people, robots can dream. But for them, ‘dreaming’ is practicing new skills in our cloud simulator.
Picture a video game with high-fidelity real-world physics built-in, where millions of simulated robots practice learning new tasks in diverse environments. That’s sim.
Sim is a cornerstone of our work, it is our accelerator for application development and machine learning model training. It can scale to create a near infinite number of environments and scenarios with high degrees of variance and randomization, reducing the time needed for the robot to learn from months to days.
Practice makes progress. We put our robots through a number of rigorous scenarios to challenge their capabilities and help them learn faster, and better.
But learning isn’t singular, it happens as a fleet, an order of magnitude larger than any one robot on its own.
Learning in the virtual world
We test and train our robots on our latest software and machine learning models in virtual world environments. We programatically set up scenes and conditions, and can rapidly reset them to a known state, accelerating bring-up, validation, training, and iteration times.
Our robots also train in staged environments. These are an introduction to real life where we can create complex scenarios quickly before deploying in everyday spaces.
Think of staged environments as robot schools, where robots can practice tasks — like sorting waste into landfill, recyclables and compost streams or wiping messy surfaces to make them cleaner — in a controlled way.
To help us in the real world, helper robots need to be in it. And they already are, supporting people in one of the places where we spend most of our time — the office.
We’re teaching robots to identify trash and correctly redistribute it, with the aim of reducing the compost & recyclable waste that ends up in landfills.
From cleaning surfaces, to monitoring air quality and inspecting spaces, we’re tasking robots with helping us stay tidy, safe, and healthy.
Our robots already assist in some Google buildings, helping to keep our offices clean and safe, all while gathering the real-world experience they need to take on more and more tasks, in more and more places.
Partnering with Google Research, we’re advancing machine learning technologies in reinforcement learning, imitation learning, multi-task learning and lidar segmentation technologies — helping our robots learn faster, and better.
We collaborate with DeepMind, pursuing breakthroughs in technology that allows robots to learn skills more efficiently from human demonstrations, and improve their own performance through trial and error.
Waymo has spent years researching lidar technology that helps self-driving vehicles understand and navigate the world. We work closely with them to push this technology even further, improving our robots and their cars in tandem.