Observation and Autonomy

In this series of articles, we will look at the history and challenges of autonomous robotics with the aim of offering insights into the missing elements of large-scale autonomy deployment.

Part 3 :

Deploying autonomy on a large scale.

After looking at the history of the autonomous car and the problem of compliance, we address here the big question of large-scale autonomous deployment.

The next industrial revolution

Industrial robotics is in full expansion in France and in the world. It allows the automation of many tasks. In France, there has been a 155% increase in robot investment over the past 10 years (2,049 units in 2010 – 5,245 units in 2019).

The automation of machines and the implementation of AGVs (Automated Guided Vehicles) in industries is one of the solutions considered in the long term to collaborate and to help people in their work. This applies transversally to several fields of activity ranging from mobility (autonomous cars) to logistics (parcel storage), via service robots in hospitals. 

Part of Amazon’s success is based on efficient warehouse automation coupled with logistics engineering that allows it to outperform many other B2C delivery services in terms of delivery time. For parcel logistics, Amazon has integrated many AGVs into their sorting centers in France. 

“The new site for AGVs, allows us to gain speed of execution and have finer granularity in sorting our parcels, organizing them by city or by area.” (Frederic Duval, Amazon France director). While this solution may increase the efficiency of Amazon’s warehouses, it doesn’t mean that its robots are smarter and that their long-term performance will be as important. The problem lies in the fact that these warehouses were built around the robot to perform their task efficiently and safely. Their ability to adapt is limited, so their evolution in this environment is also limited. 

The vision we share at Visual Behavior opposes this concept to put intelligence and adaptability back at the heart of robot activity. 

We at Visual Behavior believe that we will reach the next industrial revolution when the robotics world succeeds in integrating AGVs in warehouses without having to reorganize them.

The intended progression is not the change or adaptation of the environment but the evolution of the systems in the robots so that they adapt to the environment as it is. What prevents industries from evolving their technologies towards these new autonomous systems is the lack of alternative solutions. Most industries adapt their environments to the robots they use, not the other way around. An evolution of usage practices would therefore be just as necessary for the automation of production and logistics chains to be lighter and more flexible. 

Today, one of the major obstacles to the massive deployment of intelligent robots is the price. The use of lidars is quite widespread in small mobile robotics. The magic recipe suggested by the combination of machine learning and lidar is difficult to deploy in all sectors due to the high cost of these sensors. The companies Tesla and Comma.ai propose a cutting-edge alternative that implies that all autonomous robots can simply use cameras and artificial intelligence to interpret their environment. 

It's all a matter of interpretation

Since the 2000s, many players have entered the robotics market for individuals. Between vacuum cleaner robots, cooking robots or gardening robots, these tools are the beginning of small autonomy at affordable prices. 

The challenge of these projects, for the vacuum cleaner robot for example, can be considered from two angles :

    • The observation space : provided by a 2D location map.
    • The action space : determined by the speed of the robot and its rotation angle.

These spaces were proposed at an affordable price and, considering the relatively simple task of a vacuum robot, the observation was easy to interpret to deduce the action to take. However, this result is not reproducible on all autonomous robots. Why is it so difficult to market a gardening or cooking robot today? Because their observation space is more complex than that of a vacuum cleaner robot.

“The best mechanics or the best robotic hand is useless if the robot is not able to understand the concepts associated with a salad bowl.” (Thibault Neveu, co-founder of Visual Behavior)

Today, most computer vision systems are able to detect many objects and give them a name, but are very limited, and sometimes even devoid, of understanding their physical features and functionality.

For example, consider our development in childhood. We first learn the concept of an object like a salad bowl rather than its name. We understand that it is hollow, round, and graspable from above or with two hands, not like a sheet of paper or a chair. As we grow up, this understanding becomes what is more generally known as “common sense.” Understanding the physical characteristics of a glass or a saucepan later on will be less tedious to learn. Autonomous robotic applications need this common sense to interact with their environment and the objects in it. This same challenge of autonomy also exists for autonomous cars as they evolve in unstructured and uncontrolled environments. 

In recent years, the robotics community has developed many solutions showing that we know how to meet the challenge of action. Boston Dynamics’ humanoid robots are able to carry objects, do somersaults and even dance! Natural language processing models such as GPT-3 are able to understand complex sentences and adapt quickly to a new situation. New architectures of artificial neural networks such as transformers, as well as recent methods of reinforcement learning, allow us to consider new capacities of reasoning and action that we can endow future robots with. 



    In our view, the limitations to the large-scale deployment of autonomous robots lie in their visual intelligence and their ability to understand their environment. When we will be able to provide robots with a capacity of observation and understanding of their environment similar to human common sense, then the next industrial revolution will take place.