This week in robotics

This week in robotics

Welcome to the Bulletin by Remix Robotics, where we share a summary of the week's need-to-know robotics and automation news.

In today's email -

  • Tesla’s former head of AI spills the beans
  • Why robots are progressing so slowly
  • Argo AI’s employees are off to Ford
  • Cursed images

Snippets

Stop manipulating me - How do you manipulate heavily-occluded, unknown objects? Researchers have developed a crazy-looking “Tactile-Enabled Roller Grasper” capable of exploring objects of unknown shape through touch alone. (Research)

AI summer, Robotics winter - Chris Anderson, founder of TED asks why robotics is progressing so slowly while AI improves in leaps and bounds. His answer - There is no such thing as robotics. Drones are different from arms are different from vehicles. His advice, ignore robotics textbooks and focus on developments in cutting-edge AI for inspiration. (Opinion)

Time for Ford to focus on L3 - Ford’s CEO tweeted that he has “the greatest respect for the team at ArgoAI & what they've accomplished”. As the self-driving car company closes its door, it seems Ford will be pulling employees and resources to focus on its Level 3 Autonomy BlueCruise system. This represents a general move away from full self-driving to more modest driver assistance. (News)

Turns out people are stealing robots’ jobs - The UK is less productive than European peers in every field except finances. We discussed the UK’s poor adoption of robots but Derek Thompson believes “Britain seems trapped between a left-wing aversion to growth and a right-wing aversion to openness.”. This seems to have us digressing - “Between 2003 and 2018, the number of automatic-roller car washes (that is, robots washing your car) declined by 50 percent, while the number of hand car washes (that is, men with buckets) increased by 50 percent”. (Opinion)

Overly academic - Tired of reading jargony, over long research papers? Explainpaper removes academic waffle and simplifies sections so that it's readable by normal people… There are a whole lot of ARIVX papers we’ve been putting off… (Product)

Shameless Plug - Plug insertion has always been a huge challenge for industrial robots. Researchers have found an efficient way to train systems to plug and play using offline fine-tuning. “This kind of autonomous online finetuning bootstrapped from offline RL is far more practical than conventional RL, which requires either a huge number of trials or engineering-heavy sim-to-real solutions. I think in the future, all robotic RL systems will use offline pretraining!” (Research)

Why VCs hate us  - Why do robotics companies struggle to raise funding? Sivesh Sukumar from Balderton Capital explains why robotics companies struggle to raise VC funding even when they bring value. A few challenges -“pilot purgatory” means it can take many years to win real industrial contracts, hardware requires lots of follow-on funding to scale and finally, many companies just aren't technically feasible. (Opinion)

Or do they just love us -  AMP Robotics raises $91 million of Series C funding. The company uses computer vision and deep learning to identify and recover plastics, cardboard, paper, cans, cartons, and many other packaging types from bulk refuse. There are a fair few players in the space and its a great example of sustainable robotics. Check out our Deep-dive and list of players here. (News)

Sneaky sneaky - Turns out Google's former CEO Eric Schmidt helped write A.I. laws in Washington without publicly disclosing investments in A.I. start-ups. Though not technically illegal, some say it's unethical as he was able to influence laws that might benefit his companies over competitors… (News)

MathMETAical - Meta has created a new neural theorem solver able to solve 5x more International Math Olympiad problems than any previous AI system. This is impressive because Olympiad problems require a lot of creativity and lateral thinking. Their solution acts like a human and breaks the problem up into smaller chunks. They hope this approach could be used in automated software verification. (Research)

Video / Podcast

A great interview with deep learning expert and former head of Tesla AI Andrej Karpathy. I would skip the first 30mins unless you’re really interested in his speculation on aliens 👽. Here are some highlights -

  • The power of Transformers - Neural networks are converging to a single architecture, the Transformer. Its taken over image processing, video, and text, and is starting to make a move in robotics. Why is it so good? It’s a trainable, general computer that is highly optimisable.
  • Software 2.0 - Karpathy believes there will be a change in how we program software. We will no longer write in C or Python but instead, use neural networks to solve all software problems. This is already starting to play out with GitHub Copilot but there is still a lot to be defined and lots of opportunities for new start-ups and business models.
  • Self-driving cars -  Why is vision the best sensor for AVs? It's the highest bandwidth, and it's the sensor humans use for driving so our world is designed around it. The challenge is processing the data at scale locally in the car with low latencies and errors… This matches the challenges of deep learning we discussed last week.
  • Why Andrej could see himself coming back to Tesla in the future….

Meme of the week

We’ve turned ourselves into a meme… Remix has been discussing image generators a lot in the last two weeks and a big topic of conversation has been what new businesses will pop up as a result.

A great example is https://www.profilepicture.ai/ - send in some photos and they generate a load of arty profile pictures. They made $10k within the first 3 days of launch…Some of them are very comical.

Jack Pearson

London