Robotics: Ignore the Push for Full Autonomy and Pursue Automation as Infrastructure
There has been a minor disagreement within the mobile robotics industry. A case of definitions to be precise. Under RIA 15.08, the industry defined what counts as an autonomous mobile robot (AMR).
Those vendors whose systems can change their route dynamically count as such. Those that rely on fixed routes are instead classed as automated guided vehicles (AGVs). This might not be a particularly important distinction outside the industry, but there is a lot more marketing prestige associated with being an AMR. Because of this, some vendors are unhappy with the definitions. They argue that the lack of external infrastructure is sufficient to call a system autonomous as opposed to automated.
There is an informal acceptance in the robotics space that there is a hierarchy of quality in regard to control. Manual control is low, remote control and teleoperation are lower, but can be made better through interesting interfaces like VR. Automated and guided systems are valid and autonomous systems are the future.
Two things to discuss. Firstly, individual vendors and small robotics communities have a big say in promoting full autonomy as the end goal, and could also build other narratives. Secondly, if large-scale automation is going to deliver on the growth it promises, it may be better to retire full autonomy as the loadstar of robotics.
Human-Like Autonomy vs Infrastructure
Back to the disagreement.
Some vendors (like Fetch Robotics and Mobile Industrial Robotics) argue that only robots with the capacity to dynamically reroute after detecting obstacles can count as AMRs. In this sense, the robot is not following a virtual line in a pre-defined fashion, but figuring out how to get from A to B. It is perceiving obstacles and moving around them without any need for pre-planned routes. These systems tend to be fitted with high-end compute platforms like Nvidia ISAAC.
Other vendors, like Seegrid and Auto Guide, argue that to be an AMR, the critical point of differentiation with AGVs is not relying on external physical infrastructure. This includes hardware like fiducial markers and magnetic tapes for the robot to localize. The vendors in this predicament tend to offer vehicles with much higher payloads, where safety concerns trump the ability to reroute dynamically. When a modern automated forklift gets deployed, an engineer scans out the area with the vehicle’s sensors, often using a remote controller. When an obstacle is detected, a robot can stop or reroute across pre-defined lines, but cannot alternate dynamically. These robots are essentially infrastructure.
AGVs are the robots that have been deployed at scale and based on Amazon’s new robots, they will remain so for the largest companies. And yet, vendors fight to limit the definition of AMR because there is prestige and marketability associated with autonomy.
The focus on ‘autonomy’ is part of the more general interest in artificial intelligence. Implicit in the goal of creating autonomous robots is the belief that human-like autonomy is replicable in machines. The development of autonomous robots is thus a simulacrum of the desire to achieve complete autonomy with humans.
Based on conversations, the Robotic Industries Association’s steering committee for RIA 15.08 involved many vendors. But Fetch Robotics, a developer of dynamic AMRs, was particularly involved. This is down to a certain level of dynamic leadership. Fetch only had US$10 million in revenue in 2020, but was able to get acquired for US$300 million by Zebra Technologies, get valued at US$1 billion. Through RIA 15.08, they have influenced industry standards that give it a clear marketing advantage over its competitors. Despite being smaller and having limited case studies for their products than intelligent AGV vendors, they set the norms for what could become a very important industry.
The Factory is the Robot
The other way to see automation is as cyber-physical infrastructure, centred on the development of the outside world as opposed to the pursuit of general autonomy. This is in fact the implicit understanding of automation since the first robots were deployed in New Jersey by General Motors in 1961. While modern robots are defined by ‘sense, process and act’, their ‘thinking’ could only be human-like in the most basic sense. We are still talking about dumb machines. While the work of AI safety and ethics experts is welcomed, their horizon is markedly longer than for industrialists looking to raise productivity.
Elon Musk has discussed automation in both ways. On one level he has engaged and contributed to the pageantry of the singularity, with a lot of focus on Nick Bostrom and speculative takes on a doomsday where AI realises people are useless. This also manifests in his take on Tesla’s autonomous vehicle program. To Musk, the answer is to deploy the robot with cameras that are trained with the best possible algorithms. With enough development, there is no need for auxiliary sensors like LiDAR or for new roads of cities. In this sense, Musk agrees with the spirit of RIA.15.08.
At the same time, Musk has also been engaged in one of the most ambitious attempts to automate critical infrastructure, namely his Tesla factories. The ‘factory is the robot’ meme has circulated around tech media, and it adequately describes his ambitions. As it happens, this has proven challenging with myriad setbacks similar to the automation failures GM experienced in the eighties. But despite these problems, it is at the level of manufacturing infrastructure, as opposed to the drive for autonomous cars, where Tesla has been successful.
Although futurists are more interested in humanoid robots and self-driving, the demonstrated value of automation rests in building improved infrastructure for production and logistics. Amazon, the most successful scaler of automation in the US, has deployed 350,000 automated robots. They are guided vehicles with no dynamic rerouting capacity, thus not counting as autonomous according to RIA 15.08. While this disappoints some experts, the scalability of the solution overrides the need to deploy as much agency as possible onto one particular robot.
Even as autonomy is pursued and touted, those who are neck-deep in the industry are slowly growing the confidence to push back on the complete autonomy hype. One good demonstration of the failed proposition of autonomy is provided by the former CEO of failed self-driving start-up Starsky Robotics.1 As more algorithms are trained with ever more data sets, it becomes increasingly costly to improve them further. Much of the gains of automated trucks can be offered by remote teleoperation. There would still be truck drivers of a sort, but they could operate remotely, likely improving the job experience. Remote operation can be enabled by external infrastructure placed on the vehicle and on the road. This is a daunting prospect, but also a chance to build new cities.
There are also successful companies like RE2 Robotics which have focused on building high-value teleoperated systems, shipping 650 units so far. Instead of autonomy, they have their roadmap centred around more sophisticated supervisory control and fleet management.2 The route to success for robots in rugged and challenging environments is going to be gradual improvements in semi-automated teleoperation, rather than anything truly autonomous.
Even as mobile robots gradually acquire dynamic rerouting and appear to become autonomous, they will be reliant on virtual fleet managers, external sensors, 5G, and Wi-Fi 6 services and will need direct teleoperation capability in the case of emergencies. They will also increasingly be operating in factories, warehouses, and fulfillment centres designed around automation, and thus will be supplemented with charging infrastructure and real-time location services (RTLS) to help anchor them within their environment.
This in turn explains part of why we have many robots in factories or business parks, but very few in urban areas or in the field. We build new versions of the former, but not the latter. To get close to the aspiration for automated driving, you need new roads and new infrastructure. The current Infrastructure bill saddled in the US legislature is nowhere near ambitious enough in charting a course for deploying transformative urban technology.
Big Decisions, Small Organizations
Mobile robotics remains a very nascent market and is yet to have its path set.
There is no dominant seller, as Amazon is a self-contained company automating only its own assets, and there is as of yet no dominant buyer. The closest to one is the automotive industry, which is disjointed by geographies. The most politically adroit automakers, the Germans, are represented by the German Association of the Automotive Industry (VDA).3 Far more coordinated than American counterparts, their interoperability standard (VDA 5050) is designed to set common protocols for guided vehicles and AMRs. As non-automation players, they have got quite a bit of criticism from start-ups, including Austrian automated forklift developer AGILOX. Though VDA can impose themselves in European manufacturing, they cannot avoid robotics community pushback or alternative standards in the US or Asia.
In lieu of dominant buyers or sellers, robotics communities themselves are developing standards on their own, with RIA 15.08 being a prime example. Mass robotics, a major body for robotics companies in the Boston area, is heavily involved in setting industry-wide protocols for robot interoperability. This can extend from basic information sharing to developing common task management systems for mixed fleets. That such an underlying set of standards that could impact the whole industry is determined by a few small companies is a peculiarity of what is still a very nascent American-centred community.
It would do those in Government and policymaking well to take note of these developments, at least to the extent they do with the development of 5G standards through organizations like the GSMA.4 Automation is not just the US$50 billion in turnover we get right now, but could very well be so critical to economic performance as to demand its own government departments. Getting it wrong is going to be costly, and the European Union might well be doing just that. The upcoming AI act and machinery directive has caused a backlash among roboticists, due to new mandatory third-party certification requirements.
The EU is a robotics juggernaut, so a misstep like this spur some repositioning from the British Government. I myself contributed to a British civil service document on automation across different industries. One way to improve Britain’s weak position on automation and productivity could involve direct consultation with the representatives of RIA and Massrobotics, or even throwing the wallet at bringing some of Stanford’s best professors to Imperial.
The US Government is likewise disconnected from the robotics space. There is a real need to link up the depth of knowledge in organizations like Massrobotics and Robot Operations Group with those responsible for policy. While the general view of the tech sector is that the Government should stay in its lane, the remedy for our modern maladies might be the opposite. The Government should focus on automation at least as much and probably more than variable renewables. While solar panels are taking care of themselves and far outstripping demand expectations, projections about robots have been consistently disappointing. It is time for a cyber-physical infrastructure bill worthy of the name, with budgets comparable to the Green New Deal.
For some motivation, China has placed robotics development at the top of its priorities with its own stand-alone five-year plan.
For a non-hyperbolic, non-apocalyptic, non-utopian take on automation, I would recommend David Mindell’s ‘Our Robots, Ourselves’.5
For some excellent interviews and access to the best roboticists, check out Robot Operations Group. 6
In an act of shameless plugging, here is my 2019 essay on automation anxiety in American Affairs. 7
I have also recently done a podcast with ITIF on automation.
https://medium.com/starsky-robotics-blog/the-poor-roi-of-autonomy-f5d6f4f2dd14
https://www.resquared.com/
https://www.vda.de/en.html
https://en.wikipedia.org/wiki/GSMA
https://www.amazon.co.uk/Our-Robots-Ourselves-Robotics-Autonomy-ebook/dp/B00SI02AWK
https://robops.org/
https://americanaffairsjournal.org/2019/05/automation-anxiety-in-an-age-of-stagnation/