As soon as the news broke of the novel coronavirus we also began to hear of the importance of modeling and the phrase that would come to echo in all of our ears:
“If we only had a better model.”
My heart goes out to everyone struggling to make decisions as it seems there are no easy solutions. Most of our leaders have adopted models to help face the uncertainty. George Box explained the problem with models succinctly: “All models are wrong, but some are useful.” The recent article in the Atlantic titled “Don’t Believe the COVID Models” argues that although models are flawed and rarely predict the future, we should cling to them anyway as they help us address the worst case.[1] Having sent robots into caves, bunkers, tunnels, minefields and radioactive environments, I’ve learned the problem with depending on a static model.
What We Can Learn from Robots
Robotics is a domain where this all comes to life. Many roboticists use simulation to kickstart their projects, but when the resulting brain is placed into a physical robot, it may struggle to link its model to the events unfolding in the chaotic world all around it. The predicament our public health officials and politicians find themselves in is not dissimilar.
Rodney Brooks, co-founder of iRobot, stated that every simulation is doomed to succeed. What he meant is that within a static model, success can (and usually is) guaranteed. Once out of the model, the chaos of the real-world seeps in and everything goes haywire. The more uncertain and dynamic the situation, the more devastating is the dependence on the model. When facing the unknown, robot intelligence is characterized not by adherence to a static model, but by adaptation — context sensitive responsiveness guided by continuous feedback.
When COVID first reared its head and the phrase social distancing was used, I couldn’t help but immediately think of my early work on swarms of robots. At DARPA in the ’90s, we developed the worlds’ first 100 robot swarms for the Department of Defense and matured the technology for mapping out high radiation and finding chemical spills. Other robotics projects, focused on cognitive intelligence, depended on maps and the robots tried to reason about everything in their model. When things changed, they could not figure out what to do because their action depended on the old model. In contrast, the swarm robots didn’t use models. Instead of basing their behavior on maps, they based it on each other. They always moved forward even in the face of uncertainty, continuously adapting to the chaos all around them.
It wasn’t just the fog of battle that required adaptation. Working with a leading heavy equipment company we designed a robot forklift that could build a map of a construction yard. With many scissor lifts in constant motion, the poor forklift could rarely complete its map or guarantee a safe path. It sheltered in place, waiting for a better model. In contrast, my swarm robots were more like cockroaches. They deliberated very little, but constantly learned and adapted. For these robots, update rate trumped model-based planning.
When sending robots into dangerous, dynamic environments, I was amazed at the power of swarm intelligence. The bots learned on the fly how to interact as they helped each other adapt. Their behavior was driven largely by something we called social potential fields – not too different from social distancing. This peer to peer behavior is what biological systems do so well – ants, bees, flocks of birds, schools of mackerel, soccer players and stockbrokers. We are all at our best when we are acting with regard to our nearest neighbors, responding to the real events around us. Static models interfere with this.
Finding a Solution
So what is the alternative? It is hard to plan for the unknown. Patton used to say that plans go out the window the moment the first shot is fired. Patton opted for preparation, not planning. A lesson of the novel corona virus is that both our sense of being in control and our belief that we can successfully model our world are illusions. We should have already understood this. Edward Lorenz, the father of modern meteorology long ago explained that the flap of a butterfly’s wing in one part of the world can cause a typhoon on the other side of the planet. Unless your model has that flap, it may not be valid. Your model could have butterfly wings in it. But it doesn’t. It’s not wrong to make models, but we must view them as what they are: hypotheses based on incomplete data. The most important thing in dynamic situations such as the COVID-19 crisis is to adapt quickly. It isn’t as important to be right as it is to be prepared and then continuously update your plans based on what happens… not based on a static model.
The most promising way to do this is contact tracing based on accurate positioning. A new plan in the EU has recognized this need and is planning to deploy a large scale, never before seen effort to enable contact tracing using peer to peer positioning.[2] EU recognizes that neither GPS nor QR codes can accomplish contact tracing effectively. QR codes requires constant, deliberate use of your phone to scan codes wherever you go. GPS simply doesn’t offer the reliability or accuracy and can easily be off by 30meters in a normal city. Bluetooth communications, available on most smart phones, offers an immediate opportunity to enact peer to peer, swarm-based communication. In this model, digital handshakes are constantly sent from your phone to its neighbors to help keep track of who you have come in contact with. This has already been tried in Singapore using the TraceTogether app with encouraging results. Singapore used it to keep COVID-19 infections in the hundreds and it represents a big step towards swarm intelligence.
Unfortunately, it requires people to opt-in and the accuracy is not what we’d like it to be. Working for a prominent airport I recently analyzed the challenges associated with tracking people and vehicles using Bluetooth. It works to a point, but the accuracy is questionable. Bluetooth signals can travel up to 100 meters and the signal deflects off surfaces in unexpected ways. With a Bluetooth-centric strategy we could end up quarantining approximately two orders of magnitude more people than we really need to.
A more promising option is based on Ultra-wideband, now embedded into every new Apple phone and coming soon in Android phones as well. Ultra-Wideband radios send peer to peer messages that allow each phone to not only keep track of who it has come in contact with but also enhance GPS to within centimeters. After installing UWB into the streets of Manhattan for a Department of Transportation project we showed that we could limit positioning error to less than 10cm – ideal for autonomous vehicles, drones or contact tracing. Each phone localizes itself in reference to its neighbors, so instead of the network determining your position, your phone calculates its own position and you retain ownership of your own location data. It changes the paradigm from global to local, which is better for you not only in terms of accuracy but also improves privacy and security. UWB-based micro-positioning fits perfectly with the goal of swarm intelligence, enabling not only contact tracing but better traffic flow, ride sharing coordination and safety.
Becoming Swarm
We don’t need to become ants and bees, but we do need to embrace the principles of swarm intelligence. When considering a swarm approach, the key elements are: 1) emphasis on nearest neighbor interaction; 2) continuous, actionable feedback; 3) fast update rate; 4) context-sensitive behavior; 5) use randomness to learn and adapt.
Want to know what a swarm response to COVID-19 would look like? Consider Taiwan which was one of the earliest countries to post COVID-19 cases. It never used draconian measures or instituted a lockdown, but instead focused on adapting quickly, using contact tracing as continuous feedback. The same approach was taken in South Korea and Singapore with encouraging results. We instead chose to focus on a model based on prior data. We now know that the three biggest factors in our models – the infection rate, the death rate and the need for ventilators — were misleading.[3],[4]
Our love of model-based centralized control is causing us to lose important battles. The most important battles. According to our mind-set our sophisticated models make us smarter. According to a swarm-based understanding of intelligence, COVID -19 is smarter than us. It is more responsive and decentralized, and is no way deterred by chaos or the unpredictability of the real world. To cope, we must ourselves become “swarm.” The corona virus showed us we are all bound to our nearest neighbors. In a swarm the success of the least of us helps us all. If we weaken our neighbors by refusing them clean water, good healthcare and hygiene we find that we are all much more likely to be sick.
It is not wrong to model, but we need to ensure models are helping us prepare and respond, not fanning fear. FDR believed the only thing we have to fear is fear itself. In that light, the viral spread of anxiety may become our biggest medical concern. Instead of looking at our actual neighbors — who are mostly just walking their dogs — we are conditioned to fixate on fear. Instead we should be fixating on each other and our local communities.
We need to adapt faster. We need to move forward despite the uncertainty. There is one phrase I’ve whispered into the behavior code of all my robots: When in doubt move forward.
As CEO of W8less, Mr. Bruemmer is leading large-scale efforts to create next generation GPS for the “internet of moving things.” He co-founded 5D Robotics — which he grew for eight years into an industry leader in robotics and autonomy before it was acquired in 2018 — and worked on large scale robotics programs for the Army and Navy, the DOE, the DoT and DARPA.
[1] https://www.theatlantic.com/technology/archive/2020/04/coronavirus-models-arent-supposed-be-right/609271/
[2] https://venturebeat-com.cdn.ampproject.org/c/s/venturebeat.com/2020/04/09/can-a-coronavirus-tracking-app-be-both-effective-and-privacy-centric/amp/
[3]https://www.bmj.com/content/369/bmj.m1375?=&utm_source=adestra&utm_medium=email&utm_campaign=usage&utm_content=daily&utm_term=text