The focus on automation has sharpened in recent months, following global disruption of supply chains and labour. But what can be done now to speed the delivery of such vital solutions to help in the aftershock of the pandemic?
The COVID questions:
Throughout the crisis we’ve witnessed heartening examples of organisations pushing development limits when pressed to deliver rapid solutions. The COVID-19 ventilator challenge that many of my colleagues contributed to is an inspiring example. But this extraordinary level of productivity will be unsustainable when the global lockdown is over, and other methods may be needed to meet expectations for a ‘new normal’ of development timelines.
The impact on the way we all work is – along with wellness, retail and leisure – one of the cornerstones of our coordinated response to COVID-19 here at Cambridge Consultants. The consequences of the pandemic are far reaching. Labour forces, particularly those with seasonal workers, have seen dramatic disruption during the lockdown. Taking food harvesting as an example, one quarter of Italian food products are gathered by 370,000 foreign seasonal workers, primarily from Eastern Europe. The workforce required to meet this gap is currently heavily undersupplied.
Searching for solutions
This short supply problem is compounded by the fact that such roles will leave workers more exposed to potential infection than other remote working jobs and are less well paid. The shortfall is being helped to a certain extent for the moment by people switching from posts in restaurants and pubs. Social distancing, too, provides good reason to search for automated solutions. Taxi and Uber hire will no doubt take a long time to regain consumer trust following the lifting of restrictions, and consumer concerns will likely shift significantly towards hygiene and cleanliness.
Automation can help ease the shift, but some AI-driven solutions require many thousands of hours of training to ensure that they work effectively, thanks to ‘edge cases’ in the real world. The self-driving Uber crash of 2018 is a tragic example of this – low light, slow movement of the pedestrian and her unusual silhouette as she walked with a bicycle all contributing to the fatality. These types of edge cases are, thankfully, rare. But it is this same rarity in the training of these programmes that can make them underprepared for the uncertainty of real roads.
Mimicking the world
To address this, computer simulations called Virtual Learning Environments can be used to imitate the real world. These are training engines, capable of training real-world reinforcement learning (RL) agents entirely in simulation. For autonomous vehicles, they can take the form of detailed renders or roads to emulate physical world conditions and scenarios.
Researchers from MIT have created a photorealistic simulator called the Virtual Image Synthesis and Transformation for Autonomy (VISTA), which uses real data captured from human driving experiences to synthesise random scenarios. Due to the provenance of the data from the real world, the engine can generate more realistic situations than rendered counterparts.
It is not just information from reality that can assist with self-driving cars. To train networks faster and more cost effectively, a publicly available dataset called Playing for Benchmarks is available. It uses synthetic images from the video game Grand Theft Auto to stitch together photorealistic driving scenarios. Although this will never match the realism of the true world, a major benefit of synthetic sets is data annotation. Since everything in the game has been designed from scratch, there is a pre-existing label for all trucks, trees and pedestrians. This removes a time-costly step involved in creating the training engine.
Augmenting our abilities
Simulation of this kind can prove useful in seasonal sectors too. For farming, there are only narrow windows in the year where certain new methods can be trialled. This results in a long and slow development cycle. Virtual Learning Environments allow the circumvention of the annual cycle, which could provide considerable benefits to farming equipment and techniques.
The major challenge in agriculture simulations is uncertainty. Seed type, soil, weather variations, fertiliser, insect provenance and fungicide must all be considered, and the resulting simulation must have well understood parameters and co-dependencies.
In addition to automation, augmentation could prove valuable to supplement labour shortages for semi-structured tasks such as food harvesting. We have highlighted the difficulties associated with such tasks in previous articles. Now, a number of firms are developing solutions that can tackle them. The applications extend beyond agritech to many arenas where either social distancing or labour disruption would warrant a use case – from packing shopping bags to providing home meals delivery.
If robotics can take care of the most obvious cases in these tasks then humans can achieve more with less, reserving the more complex edge case tasks for ourselves. This would enable jobs to be more intellectually stimulating, increase productivity and could assist in bridging the potential labour gap.
Join the debate
While social distancing is here to stay, automation and augmentation could enable humans to do more with less, reducing human-to-human contact. However, enhanced simulation will be required to get it to a usable state in short order. The tasks that automation can take on will need to be carefully selected to contain the right level of uncertainty for these systems.
At Cambridge Consultants, we are continuing to reach out to clients and colleagues to understand what the post-COVID world will require, and what technology will be necessary to support it. Please drop me an email if you’d like to discuss any of these topics further.