This Week in Robotics 05/05

Welcome to This Week in Robotics, where we provide weekly insight into robotics, autonomy and AI.?

This week -

• Football robots
• A win for self-driving cars?
• The end of big AI models
• News from Logimat?
• A jihad against AI

Research

DeepMind trains humanoid robots into football players.?

How they did it -?

The agents were trained in simulation using deep reinforcement learning and skills were transferred to real robots "zero-shot" -

• First, they identified the individual skills required to play football like - running, kicking, and getting up when knocked over. They initially trained the bot to excel in these isolated skills
• Without this initial decomposition, the agents find a local minimum and end up rolling on the ground with the ball rather than walking and kicking. This works but it's not pretty
• Once the individual skills were perfected they composed the skills together and trained the robots to play end-to-end football
• The training rewarded 9 behaviours including - ball velocity, player velocity, uprightness and of course goal scoring. From these simple incentives basic football strategy emerged.

Limitations?

The robots did not use onboard sensors but instead relied on external motion capture systems to give them knowledge of the ball, goal, and opponent.?

They did try using onboard cameras but this led to worse results and a larger gap between simulation and reality. The limited field of view rendered the environment only partially observable, forcing the agent to do more complex computations to deal with the uncertainty.?

Graph

F-prime has released a report analyzing the state of robotics investment.

News

Cruise self-driving taxis now operate 24 hours a day, seven days a week across all of San Francisco. Only employees will have access to the whole region for these hours but "power users" will also have access to daytime rides across the city. This is real progress and comes at just the right time. It's been a brutal year for the industry, having seen a number of high-profile bankruptcies and a 58% reduction in investment since 2021.?

Last week we visited Logimat, a logistics trade fair in Germany that hosts over?60,000 people. A few takeaways from the event -

• Bots are booming - An entire hall was filled with 10s of mobile robot companies. What makes this interesting is that most (but not all) of these companies had very little differentiation. They often share the same form factor, the same hardware and the same use cases.?
• Shifting from growth to profit - Many robotics OEMs stated an interest in cost saving as profit becomes a growing driver. Expect to see lots of consolidation in the next few years, see the next snippet.
• Avoiding Collision Avoidance - ?Mobile robot OEMs advise customers to avoid using collision avoidance as it slows down production and reduces throughput. Instead, they recommend dedicated robot zones. This works well for purpose-built facilities but is challenging for mixed-use environments. Maybe better collision avoidance software is required?

6 River Systems, a mobile robot company is being sold by Shopify to Ocado. This comes only 3ish years after Shopify made the acquisition for $450M. Unfortunately, Shopify seems to be struggling post-COVID, they plan to lay off a further 20% of their workforce and have sold their logistics arm to Flexport.?

“We [Google] Have No Moat, And Neither Does OpenAI”. A leaked internal document from a Google employee has suggested that the barriers to developing state-of-the-art AI are dropping so quickly that big tech can't keep up with open-source development.

It's just one unknown employee's opinion but it's pretty damning, some quotes -?

• "We have no secret sauce. Our best hope is to learn from and collaborate with what others are doing outside Google. We should prioritize enabling 3P integrations."
• “Open-source models are faster, more customizable, more private, and pound-for-pound more capable.”?
• “They are doing things with $100 and 13B params that we struggle with at $10M and 540B. And they are doing so in weeks, not months.”

It's worth reading the section on LORA, a method that allows for cheap and easy fine-tuning of large language models making them a lot less large, which has ”solved the scaling problem”.

OpenAi’s CEO Sam Altman agrees that the age of scale is over.

“I think we're at the end of the era where it's going to be these, like, giant, giant models…. We'll make them better in other ways.”

Opinion

Michael Cuenco calls for a jihad against AI - “The more generalized, generative, open-ended, and human-like the AI, the less permissible it ought to be”.?

This sentiment is becoming more and more common with “AI Doomers”. Eliezer Yudkowsky discussed the nuking GPU clusters and Geoff Hinton left Google to much fanfare.

The most interesting aspect of Cuenco’s article is his reference to pop culture which highlights how long this fear has existed.?

Cuenco was inspired by the Dune series which features “the Butlerian Jihad,” a crusade in which humans rise up against “thinking machines”. In the books, this is more than a ban but a full religious movement based on the principal “Thou shalt not make a machine in the likeness of a human mind.”

Dune was written in the 70s but the inspiration for the Butlerian Jihad is even older. It derives from the work of the Victorian writer Samuel Butler, whose 1863 essay “Darwin Among the Machines” and novel Erewhon, where humanity confronts machines smart enough to replace their creators, leaving only two possible outcomes: the destruction either of humanity or of the machines.

As Butler saw it, there could be no partnership between humans and human-like machines. As machines become more human and we outsource more and more to them - we progressively dilute our human acumen and creativity while enhancing theirs.?See the Tweet below for a counter argument.

Tweet