Shared Autonomy for Collaborative Manufacturing: The Future of Interactive Robotics

Advances in Robotics in recent years have seen the creation of ever more increasingly powerful and compact mechanical devices through breakthrough advances in high fidelity actuation systems, low price and extremely rich sensing modalities (like stereo cameras, depth sensors (RGB-D), Lidars, IMUs) and significantly enhanced compute power in compact packages.

However, great challenges remain in the operations of these robots in domains outside the well-controlled laboratory situations. So, for safety, solution has been to either move very slowly or isolate from humans. Both are impractical for real world applications of robotics except in very structured manufacturing scenarios.

So what is the future?

At the University of Edinburgh as well as several other leading robotics labs around the world, we have been developing novel compliant technology (actuators and controllers capable of reacting in a safe manner while interacting with humans) and dynamic, reactive planning that helps solve these problems in real time; allowing robotic technology to be more accessible to much closer human robot interaction scenarios like co-workers in agile manufacturing, health-care, robotic surgery, and exo-skeletons for human assist systems.

Many robotic applications we see in manufacturing and remote handling (nuclear), space or surgery use the concept of teleoperation this involves the robot precisely following a sequence of movements that are kinaesthetically demonstrated through some user interface. Recently, I saw first-hand, the extreme skilful teleoperation of robots in a fun setting: I am one of the judges on the recently rebooted series of BBC Robot Wars an extremely popular technology show with a fanatical following.  There were some amazing, agile and destructive machines and some terrific young drivers.

I also noticed that if I were given the winning machine and asked to drive it against an experienced competitor, I stood no chance. Teleoperation, in this and many other more practical settings, heavily relies on the skills of the operator and this can be a significant barrier to more widespread adoption of robotics in manufacturing due to the skill gap - in other words, not everybody can effectively operate the robots, even if the engineering design is amazing.

On the other extreme, many robotics research agendas are striving to develop technologies for 'fully autonomous' robot operations - 'look ma, no hands'. Many world leading groups, including our own, have been working on this problem. These are very relevant is many domains such as underwater AUV navigation, remote operations in communication denied or corrupted environments. Testing and validating autonomous behaviour in real world environments is a significant challenge. Recently Professor Rodney Brooks (of iRobot and Rethink Robotics) and Professor Marc Raibert (of Boston Dynamics) were quoted as saying that full autonomy is very far away in spite of impressive videos like that of the Boston Dynamics robots making the rounds in the news.

Full autonomy has several components including being able to define an 'optimality' or 'goodness criterion' which is intuitive and doable in many cases but in addition, the robot must define and generate intentions. Generating intent and testing whether they result in an expected behaviour can be extremely hard for complex robotic platforms. Google DeepMind recently showed that by simulating the environment and playing simulated games against its own various versions, it can solve complex AI problems like the game of Go.

However, it had to rely on lots of examples from real games. Trying to do these on real robots is quite simply impossible - due to the complexities and the variants that myriad of action possibilities can generate. Most importantly, it is hard to simulate the reality to an extent that we can guarantee success every time - something policymakers and legislators are quite keen on ensuring.

So, how does it look from the perspective of creating a business model for robotics in agile manufacturing? Jerry Pratt, a Professor at iHMC in Florida was quoted as saying, 'any business typically needs a MVP (Minimally Viable Product) and this is very hard in robotics since the simple, cheap robots are no good (as of now, at least!) and will most likely lose you money and so there is no gradual scaling possible with market uptake' - the concept of graceful degradation with level of investment is not quite true in this field. This is illustrative of the significant research challenges that still exist in the control, sensing and autonomous planning and navigation on complex robotic platforms - still active areas of research in many universities including my group.

So how can we make the benefits of robotics accessible to us without worrying about the skill gap (teleop scenario) or waiting to reliably solve the 'full autonomy' problem? We can think of a concept that is perhaps deployable now and also perhaps socially much more acceptable. This is a paradigm that marries the best of both worlds: precision and accuracy of robots and contextual decision making capabilities of the humans one I would like to call 'Shared Autonomy'. This is a paradigm that devolves significant autonomy to the machine while still having humans in the loop as key decision makers.

The key research challenge is to create systems that have the ability to seamlessly change the levels of autonomy depending on the task and even within a task at different moments in time. The main emphasis now shifts to how to reduce the cognitive load on the person in charge of a task and how robots can make some of these task 'idiot proof' in the presence of significant delays, noise and communication challenges.

An example of an exciting project we are working on involves an ambitious 'out of this world' mission. NASA in collaboration with the University of Edinburgh is working on a Valkyrie robot. The NASA Valkyrie is a humanoid robot, about 170cm tall, weighing around 120kg, and has very similar levels of dexterity as a human being. It is equipped with a range of sensors including stereo cameras, touch sensors, torque sensors and a LIDAR a kind of scanning laser. All of these are used to sense the world in depth, colour and forces.

It is bipedal capable of walking on two legs, which makes it very agile as well challenging to control. All its joints are driven with very powerful electric motors and there is an on-board high density battery and three state of the art computers making it capable of fully autonomous behaviours. The computers on board are used to receive all the sensor information, process it and make decisions about the next step the robot should take sending commands to the individual joints. This robot is one few humanoids in the world that is fully torque controlled. It is one of the futuristic project of NASA where it  plans to use humanoid robots to send un-manned robotic missions to Mars.

The idea is that it would make manned missions much cheaper, safer and efficient if we could use humanoid robots to pre-deploy essential capabilities on the surface of Mars ahead of the manned mission.

These examples and narrative also addresses the issue of heavily topical: 'Will robots take my job?' question. I am personally of the opinion that robotics will be more economically viable and much more socially acceptable operating within the 'Shared Control' paradigm. They are very much designed to be co-workers, taking over dull, dirty and mundane things that are typically easy to automate, and require much less 'cognitive' prowess while freeing up time and resources. Perhaps the more pressing and important question is the distribution of this potential wealth (and benefits) that a robotics revolution will bring.

This question is not very different from the issue of identifying essential skills for the next generation and planning ahead to equip people with these skills and ensuring a rounded education that match the demands and the skill gap. Also, security of your data and privacy legislations should be carefully looked at personal data is the 'new oil', since several of these advances will require you to divulge various levels of user generated data in return for potential huge benefits. And finally, there is the issue of responsibility and legislation laws that needs to be addressed when autonomy is devolved - and to create and develop systems that work closely with these legislations such that clear accountability can be established.

An exciting and transformational future for innovative, agile and collaborative manufacturing beckons and undoubtedly, interactive robotics and artificial intelligence are at the heart of it!