Radical Reads

Radical Investment in AI & Robotics: Logistics AI Startup Covariant Reaps $40 Million in Funding Round  (Wall Street Journal)

“Artificial-intelligence robotics startup Covariant raised $40 million to expand its logistics automation technology to new industries and ramp up hiring, the company said Wednesday.

The Berkeley, Calif.-based company makes AI software that it says helps warehouse robots pick objects at a faster rate than human workers, with a roughly 95% accuracy rate. Covariant is working with Austrian logistics-automation company Knapp AG and the robotics business of Swiss industrial conglomerate ABB Ltd., which provides hardware such as robot arms or conveyor belts to pair with the startup’s technology platform.”

Radical Commentary: We are very excited to announce our investment in Covariant, a universal AI brain for robotics based in Berkeley, California. The founding team includes world-renowned UC Berkeley Professor Pieter Abbeel, and PhDs from his lab including Peter Chen (CEO) and Rocky Duan (CTO).

We originally met Pieter over three years ago when we invited him to Toronto to speak at the Machine Learning Advances & Applications seminar series held at the Fields Institute for Mathematics. Tomi and Jordan had co-founded the speaking series with Professor Richard Zemel while creating the Vector Institute (the series is now run by the Vector Institute). Pieter is widely regarded as the world leader in applying AI to robotics, a field that by definition integrates theoretical and applied work. We followed his work and the creation of Covariant with keen interest.

When the company launched out of stealth in January with a NY Times feature and lots of other press, we immediately began a comprehensive outreach to the founders directly and through mutual friends and their existing investors in anticipation of a funding round. As part of our due diligence, Jordan called Geoffrey Hinton, who was an angel investor in Covariant. Geoff explained their discussion in a tweet this week: https://twitter.com/geoffreyhinton/status/1258143109344624643?s=21

As we have developed our investment thesis in one industry after another, it has become clear that AI has the potential to dramatically advance the use of robotics by turning robots from dumb automotons into learning machines. One consequence of COVID-19 is the need for hardening up supply chains in everything from agriculture to pick/pack/ship warehouses through the use of more robots to reduce reliance on humans for repetitive work in dense environments.

Why do we think Covariant is the solution? First, the product is software that is able to be deployed with both new and existing robots without requiring their replacement. That opens a much bigger growth opportunity than integrated software + hardware solutions that require replacement of existing robots. Second, the product is hardware agnostic. Covariant partners with existing robot manufacturers like ABB and deployment companies like Knapp which bundle the Covariant software with hardware they have already deployed or will deploy in future. This starts with what the company calls ‘the market for hands’, which is a trillion dollar annual market. Third, Covariant’s technology is far ahead of the competition. This Fortune article explains how ABB developed a competition to benchmark AI robotics software companies. None of the other 19 competitors completed the competition in 48 hours, while Covariant completed it in 3 hours with high accuracy.

Covariant uses deep imitation learning, deep reinforcement learning and meta-learning to automate manual tasks performed by human hands, starting in supply chains. Like other smartly designed AI systems, one of the key advantages of the Covariant product is that its AI learns from every one of its deployments. As the number of deployments increase, the resulting ‘network effect’ means the centralized ‘Covariant Brain’ AI gets better everywhere. The result is a constantly accelerating lead over any competitors and an increasingly wide technology and product moat.

AI Advancements: AI and Efficiency  (Open AI)

“..since 2012 the amount of compute needed to train a neural net to the same performance on ImageNet1 classification has been decreasing by a factor of 2 every 16 months. Compared to 2012, it now takes 44 times less compute to train a neural network to the level of AlexNet2 (by contrast, Moore’s Law3 would yield an 11x cost improvement over this period). Our results suggest that for AI tasks with high levels of recent investment, algorithmic progress has yielded more gains than classical hardware efficiency.”

Radical Commentary: OpenAI has found that computational costs to train AI systems in vision and language translation tasks have substantially fallen since 2012. Computational costs are directly linked to the efficiency of the AI algorithms and the hardware used. The improvements in algorithms over the past eight years is the key reason for the overall decline in computational costs, rather than improvements in hardware.

The article discusses the need for a generally accepted benchmark to measure advancements in the efficiency of algorithms, in a similar fashion to how the efficiency of hardware is benchmarked against the number of transistors in a dense integrated circuit.

The article also sets out some limitations to the findings. For instance, the conclusion on the efficiency of AI algorithms is based on a few data points across a few narrow tasks. Second, OpenAI does not extrapolate on the findings as it is too early to speculate on future advancements in algorithm efficiency with only eight years of data. Meanwhile, we are also seeing startups working to improve the efficiency of hardware specifically for AI training and inference purposes, including our portfolio company Untether AI. The combination of algorithms still being in their infancy, combined with potential disruption in hardware innovation makes forecasting difficult.

COVID & Masks: Paris Tests Face-Mask Recognition Software on Metro Riders  (Bloomberg)

“The Paris metro authority is testing CCTV software to detect whether travelers are wearing face masks.

It’s part of the city’s efforts to end lockdown and help prevent the spread of Covid-19, but it’s raised concerns from the government’s privacy watchdog.

The authority began a three-month test of its software from French tech company Datakalab this week at the Chatelet-Les-Halles station in the heart of Paris, normally used by about 33 million passengers per year. Monitors will have access to a dashboard with the proportion of riders believed to be wearing masks.”

Radical Commentary: This kind of facial recognition software is already used widely in Asia, but as we have commented before, the COVID crisis is accelerating the adoption of technologies that would normally be subject to long review and approval processed, and social acceptance in western countries. We believe that it is important to have quasi-governmental organizations in place which can nimbly review these technologies and the implications of their widespread adoption before that happens rather than afterward.

Cities: Sidewalk Labs Announces it will “No Longer Pursue” Quayside Project  (BetaKit)

“Sidewalk Labs has announced it will no longer pursue its project at Quayside in Toronto. The company noted the decision was due to the current “unprecedented economic uncertainty.” Sidewalk Labs and Google will both remain in Toronto.

The company, a subsidiary of Google parent company Alphabet, has worked on its smart city proposal for the Quayside neighbourhood over the last two-and-a-half years, alongside Waterfront Toronto, the tri-government agency overseeing the development.”

Radical Commentary: Sidewalk Labs has ended its partnership with Toronto to build a smart city in an undeveloped area adjacent to the downtown core. While the project was beset by controversy over scope, privacy and other issues, it is unfortunate that it won’t proceed in Toronto. This kind of global ambition is important for technology ecosystem building. Happily, we are seeing that level of ambition across Canada and in particular in Toronto/Waterloo.

For a view of what the future might look like for AI in cities and transportation, we recommend reading about the “AI City Challenge”, originally launched in 2017 by NVIDIA and now co-organized by a number of research universities. 315 research teams participated in the challenge this year, which benchmarks capabilities of computer vision systems applied to transportation systems in four categories:

1. Multi-class, multi-movement vehicle counting.
2. Vehicle re-identification from different camera perspectives.
3. City-scale multi-target multi-camera vehicle tracking.
4. Traffic anomaly detection.

Research teams from BAIDU won in three of the four categories, perhaps indicating that these technologies are developing fastest in jurisdictions where State and tech companies work together.

Navigating Crisis: A Radical Talk with Ed Clark

Ed Clark was the CEO of TD for 12 years and is widely credited with turning the institution into a retail banking powerhouse while successfully navigating the turmoil of the 2008 financial crisis while preserving a AAA rating and not taking any government money. Today he is the Chair of the Vector Institute for Artificial Intelligence and a Partner at Radical Ventures.

Ed sat down with Jordan Jacobs, Co-Founder and Managing Partner of Radical Ventures, to discuss the lessons he learned from 2008 and to share his thoughts on how businesses and governments can navigate this crisis. He also touched on the important role technology and AI will play in the economy that emerges.