This Week in Robotics 26.11

Welcome to the Robot Remix, where we summarise the week's need-to-know robotics and automation news.

In today's email -

• How ants solve the traveling salesman problem
• Dance robots
• Meta’s AI wants to rule the world
• Humanoid robots will be worth billions

News

License to kill - The San Francisco police department has proposed using robots with “deadly force”. They currently use bots for area inspection and bomb disposal but the department wants to expand their use for criminal apprehensions, critical incidents, executing warrants, or during suspicious device assessments. The SFPD proposal would allow these robots to kill people “when the risk of loss of life to members of the public or officers is imminent and outweighs any other force option available to SFPD”. We’re living in a brave new world.

Doctor Doctor - Rishi Sunak urged the NHS to innovate and improve productivity by embracing robots. He spoke at a recent conference in Birmingham, saying - “…with more robotics and automation, we can drive up productivity. It reduces some of the pressure on labour, and creates good jobs for people,”. Preach.

Leaving Arrival - Denis Sverdlov, the Russian billionaire and founder of Arrival has stepped aside as CEO. This comes shortly after a strategy U-turn that shifted the car company's focus from the UK to the US. Arrival’s valuation has dropped by more than 95% since its peak of $15bn. On Thursday it was worth only $215m. Pair with -  the pitch deck Arrival used to raise $629M.

Amazon Alexa is a “colossal failure” - It's on pace to lose $10 billion this year. The BI report spoke with "a dozen current and former employees on the company's hardware team," who described "a division in crisis." Just about every plan to monetize Alexa has failed, with one former employee calling Alexa "a colossal failure of imagination," and "a wasted opportunity.” Amazon has significantly reduced its headcount which has led to "the swift downfall of the voice assistant and Amazon's larger hardware division.” This is interesting to compare to the hype Large Langauge models are garnering. Is finding a viable business model challenging, are LLMs ready for use in the wild or is Amazon behind the curve?

Project Pitchfork - Alphabet has a secret project to develop code-writing AIs. The project known as Pitchfork (ominous) is a spin-out of X, their research / “moonshot” division. The tool is designed to learn programming styles and write new code based on learnings. It seems like Alphabet is trying to play catch up as this represents a larger move into generative artificial intelligence.

Recruitment is cooling down - Deepmind has frozen intern recruitment with some applicants finding out just before their interviews. Here’s a useful thread of AI companies still hiring.

Research

Autobots Assemble - Researchers at MIT’s Center for Bits and Atoms (CBA) have developed a swarm of modular robots capable of self-assembly. Each robot is the volumetric equivalent of a two-dimensional pixel or Voxel which can move, and transmit data and power. These Voxels can assemble themselves recursively, making more robots and hierarchically, making larger robots. Why is this interesting? I haven't seen a compelling use case yet but they look cool…

World domination - Meta is on a real streak at the moment. Last week everyone was talking about Galactica, this week it's all about Cicero an AI system that beats 90% of human players in the game Diplomacy. Why is this interesting? Diplomacy is a multiplayer game best described as a mix of Risk, poker, and Survivor. It was JFK’s favourite game, involving both collaboration and backstabbing. To succeed the AI must have the ability to reason about the beliefs, goals, and intentions of others and an ability to persuade and build relationships through dialogue. Should we worry that Meta has built an AI that can manipulate people to achieve the goal of world domination? What else is interesting? Cicero’s has a complex and interictally designed structure. It did not emerge spontaneously from the basic data but was crafted with a lot of thought from its developers. Its success stands in contrast to scale maximalists who believe that the best results for AI come from general algorithms and petabytes of data.

Teamwork makes the dream work - Researchers have developed a framework to help humans and robots collaborate through the division of labour. Partially autonomous robots are tasked with handling the majority of decision-making where they work reliably; meanwhile, human operators monitor the process and intervene in challenging situations.  Why is this interesting? In chess “Centaurs”, human / AI partnerships beat both solos AI and humans. It makes sense that this approach transfer to other areas. Also, data generation is very tough in industry and this is a great way to bridge the gap while still bringing value.

Robot dance moves - Researchers have developed EDGE. Editable Dance GEneration,  “a state-of-the-art method for editable dance generation that is capable of creating realistic, physically-plausible dances while remaining faithful to the input music”.  The model evaluates dance quality generated by the system using quantitative metrics on physical plausibility, beat alignment, and diversity benchmarks… Why is this Interesting? In case we want to bust out coordinated dance moves with our robot servants (see Ex Machina for reference)…

Investment

Soft Robotics Inc. secures $26M led by Tyson Ventures - The company designs and builds automated picking solutions using proprietary soft robotic grippers, 3D machine perception, and artificial intelligence. Their product allows high-speed picking of products with challenging shapes, variable sizes or soft forms.

Opinions

Worth their weight in gold - Goldman Sachs estimates that the humanoid robot market will likely be worth between $6 and 154 billion by 2035. A market that size could fill up to 126% of the labor gap, and as much as 53% of the elderly caregiver gap. A lot needs to happen first -

• Their battery life/charging speed needs to improve to the point where one hour of charging produces 4 hours of work
• Lots of advances are required in mobility, agility, computing, sensing, learning, etc. GS is very vague about this and it seems like a big hole in their assumption…
• Production costs need to drop by 15-20% / year so that humanoid robots have a payback period of 2 years.

Video

A beautiful video by Sebastian Lague investigating how ants approach the traveling salesman problem. The real value in this video is seeing how Sebastian uses code to better understand complex phenomena - he starts small and iteratively improves as he learns. The results are deeply aesthetic.

Tweet

Many our doubting this rumour but the hype for GPT4 is mounting.