This Week in Robotics 19.11

This Week in Robotics 19.11

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 -

  • Scary GPT-4 memes
  • American Dynamism and its dissenters
  • Research into helping robots to walk on rough ground,
  • How to manipulate objects in complex ways


It's a dog-eat-dog world - Boston Dynamics is suing Ghost over similarities in its robot dog. Boston Dynamics has called on Ghost to suspend the manufacture of its quadruped robot over several alleged patent violations including the ability to self-right and climb stairs. Ghost has since fired back - “Rather than compete on a level playing field, the company chose to file an obstructive and baseless lawsuit... in an attempt to halt the newcomer’s progress. Boston Dynamics is drawing on their considerably larger resources to litigate instead of innovate.”.

Move too fast and you'll break things - Earlier this week Meta released Galactica, a Large Language Model (LLM) for science that can summarize academic literature, solve math problems, generate Wiki articles, write scientific code, etc. The whole hype cycle happened in a few days, people loved it, then they hated it and now it’s been pulled from public use… there is such a thing as shipping too early.

Self-driving slows down (again)- Volkswagen has decided to delay plans for its flagship EV factory from 2026 to 2030. It was only March when Volkswagen announced that it would invest 2 billion euros in the new German factory for manufacturing self-driving, long-range electric EVs

Controversial - Israel has deployed remote-controlled robotic guns in West Bank that can fire tear gas, stun grenades, and sponge-tipped bullets. This is another step towards weaponized robots, a much greater concern than other AI existential risks (see Opinion for more).


Pickle raises $26M - Pickle Robot Company, which offers automated robotic systems for unloading trucks, has raised a $26 million Series A. “We are teaching off-the-shelf robot arms how to pick up boxes and play Tetris.”

DHL Supply Chain invests $150 million in warehouse robotics - This is the largest investment in robotics and automation DHL has made in Asia-Pacific, with 1,000 robots to be deployed by 2025. What’s on the menu? Assisted picking robots, Goods-to-Person vehicles, automated inventory management, point-to-point picking robots


Get a grip - Researchers have demonstrated that simple grippers can be used to undertake complex manipulations. The team used Reinforcement Learning to enable a robot arm to find creative strategies for enhanced 2-finger gripper dexterity. Why is this useful? Complex grippers are expensive, inaccurate, and over-engineered for industrial environments. Finding creative ways to use robust and industrially tested technology brings down costs while increasing capabilities.

A step in the right direction - Researchers have presented an end-to-end locomotion system capable of moving through rough terrain using vision alone. The robot dog is able to navigate stairs, curbs, stepping stones, and gaps using a single front-facing depth camera. Why is this useful? Traditionally this challenge requires expensive Lidar sensors and is susceptible to failure - another step to simplify robotics hardware. Lots of videos with this paper, including failures so +1 from us…

They have it in hand - MIT has developed an automated reorientation system. Why is this useful? In-hand manipulation is challenging for automated systems as it requires the ability to keep hold of a part while moving it – which involves lots of dexterity and an accurate prediction of how best to actuate a specific product. The MITs system was trained offline using Offline Reinforcement Learning. Also cool – both hardware and software are open-source.

Teaching robots uncertainty –  In robotic piece-picking, a 3D bounding box is used to ensure the arm accounts for clearances needed to pick, move, and place objects without collisions. The flaw in this approach is that vision systems often have incomplete information leading to failure and damaged parts/robots. Covariant has taught its systems to know what it doesn’t know - using "autoregressive probabilistic modeling”. If a part is occluded, tightly packed, or not fully visible the robot will take into account its level of confidence in the bounding box. If it has low confidence, the robots know to move slowly and provide extra clearance to avoid collisions.


Rose-tinted glasses - Scott Belsky of Adobe explains that we’re getting a bit ahead of ourselves on the impact of generative AI’s disruption capabilities. Demos may be impressive but he asks us to question “Magic Boxes”. Is the cost structure too subsidized to make the business viable? Are the moats real? Is the final mile for commercial use longer than advertised? He acknowledges the impact of generative AI will be large but believes that “the role of human curators and the story behind a creation will be more important, not less.“ as “human stories outperform pixels + brush strokes when it comes to emotional response”.

Don't worry about AGI- Joshua Achiam of OpenAI lays out his thoughts on the existential risk of super-intelligent AIs. “It’s real, but I think probabilities of doom by AGI instrumentally opting to kill us are greatly exaggerated". In his opinion, we have over a decade before we need to worry… What we should be concerned about is that "the world gets increasingly weird and hard to understand because AI systems run lots of things in ways we can't explain, creating correlated risk”.


Another interesting podcast with Hadrian's CEO, Chris Powers. Hadrian is the current poster child of a16z’s American Dynamism investment thesis. the pitch is that -

Advanced manufacturing is critically important to a thriving country and the US (and the UK) has let its industry ossify and decay to a dangerous point. The key challenges -

  • Aerospace, defense and space primes rely on small contract shops run by an aging population
  • There is a labor problem as young people don’t want to work in manufacturing
  • Critical knowledge is lost as key employees retire
  • Incentives don't reward innovators and budget/timelines constantly overrun

The solution –

  • Hadrian is becoming the most advanced contract shop for Tier 1 aerospace & defense companies
  • They bring together experience machinists and software engineers. They automate any tasks that can be automated –  from design to machine tending.
  • Where technology isn’t robust enough for the industry they codify knowledge and create easy-to-use tools to simplify the process and ease onboarding

A few points I really like about Hadrian –

  • Incentives - Chris flags that manufacturing knowledge isn't shared due to tribalism - in traditional manufacturing, sharing your skills might give someone else an edge. At Hadrian, machinists are given as much salary and equity comp as software devs ensuring it's in their interest for the company to succeed.
  • Dashboard & clarity - as their production process is so tightly controlled they can provide clients with an accurate dashboard of their part as they go through production. This is huge - we’re constantly chasing manufacturers for late components and getting a straight answer is not easy.

There is a lot to like here but pair it with the tweet below for a different take…


Things are getting spicy in the American Dynamism vs Crypto competition….


The rumors around GPT-4 are getting silly/scary and it's not being helped by OpenAI’s CEO dropping threatening memes.

Our favorite take on GPT-4 is that it naturally outputs at the same level of quality as the cherry-picked outputs on Twitter. Which is still huge

Jack Pearson