Groot is good, IROS wants demos, The Future of Robotics in SF on June 27
plus some great deep reads so keep scrolling
Last week I asked you to submit to the Office of Science and Technology Policy Call for Information about National AI Strategy. Turns out that many people are frustrated trying to find out just how and where to submit. And so the direct and correct link to submit (for OSTP AI Strategy Call) is HERE.
There are a few more government agencies also looking for input into the AI/Robotics National Strategy. NTIA (submissions just closed), NIST (can’t find it because the referenced page is ‘down for maintenance’). Do you know of any others? Meanwhile there are 3 meetings coming up (June 20, 22 & 27) of the National AI Advisory Committee if you really want to policy nerd out. See you there!
Big thank you to everyone who has become a paid subscriber - it makes you an angel investor in the work I do at Silicon Valley Robotics and as a startup advisor, and accelerator. And maybe we can do more to share your experience and expertize with the robotics startup community. Mentor Office Hours, maybe startup profiles. We’ll keep exploring.
Quick Reads:
Disney Imagineering let Project Kiwi loose! Groot is now free-roaming at Disneyland.
A robotic baby Groot will soon be roaming Disney parks - Interesting Engineering
Ikea trembles as scientists invent flat packable robot - The Byte
Kodiak Robotics Don Burnette discusses how autonomous technologies can transform existing commercial and military vehicles into uncrewed ones - Defense & Aerospace Technology Report
Cruise teams up with Replate to launch national food rescue partnership - Cruise
Trabotyx field robot with customized mechanism for weed control - Future Farming
Harxon: precision GNSS solution for robotic lawn mowers - Future Farming
United Airlines teams up with Eve Mobility to launch eVTOLs in San Francisco - Interesting Engineering
NVIDIA Research Wins Autonomous Driving Challenge, Innovation Award at CVPR - NVIDIA
Manufacturing is sputtering in the world’s largest economies - CNN
German robotics industry to hit an all-time high in 2023 - VDMA & The Robot Report
World’s most remote robot automates Amazon reforestation project - Robotics Tomorrow
Robot Passes Turing Test for Polyculture Gardening - IEEE Spectrum
Researchers create new robotic bee with full freedom of movement - Interesting Engineering
At last you can play tennis with a robot - Mashable
Raspberry Pi Quadruped Provides Low-Cost Research Solution - tom’s Hardware
AFWERX wants to build out autonomy ‘proving ground’ for testing new tech - Breaking Defense
Hyundai bets big on future mobility - The Investor
Researchers warn of ‘model collapse’ as AI trains on AI-generated content - VentureBeat
Rethinking AI benchmarks: A new paper challenges the status quo of evaluating artificial intelligence - VentureBeat
EU votes on first-ever regulations for safe and transparent AI - Interesting Engineering
A Former Pilot on why Autonomous Vehicles Are So Risky - IEEE Spectrum
Improving Autonomous Robots: dataset from UCB & Toyota - Center for Data Innovation
Aurora Releases Open-Source Autonomous Driving Dataset - Aurora
Autonomous vehicle takes part in the legendary Mille Miglia - EEnewseurope
ISEE wins “Autonomous Truck of the Year” in 2023 SupplyTech Breakthrough Awards - BusinessWire
Fernride nudges yard trucks toward full autonomy with $31m in new funds - TechCrunch
Annual Industry Report Series: Technology Innovation - MHIview
IERA Award Recognizes Innovative Startup at ICRA 2023 - IEEE RAS
Anh is 10 and designing robots. Educators hope more girls will join her - SBS News
Robotics Events:
June 19-22 - TechConnect World in Washington DC - TechConnect Org
June 20 - Revealing the AI 100: The World’s Most Promising Companies - CBInsights
June 20-22 - FITMA in Mexico City - Feria Internationale
June 20-22 - Sensors Converge in Santa Clara - Sensors Converge
June 26-29 - Collision in Toronto - Web Summit
June 27-30 - Automatica in Munich - VDMA
June 27 - The Future with Robots - Silicon Valley Robotics and IEEE RAS at the Commonwealth Club SF
July 11-12 - VB Transform: Get Ahead of the Generative AI Revolution - VentureBeat
Sep 19-21 - FIRA USA Salinas - FIRA
Sep 28 - The Autonomous Systems Showcase - Deep Tech Showcase
Oct 1-5 - IROS in Detroit - IEEE RAS
IROS is calling for demos - and so is Hello Robot.
Dear Hello Robot community,
You might have seen that IROS 2023 recently released a Call for Demonstrations for their Demo Expo taking place at the conference on October 4th:
“IROS 2023 is piloting a Demonstrations Exposition that will provide researchers focused on physical demonstration the opportunity to showcase their hardware during an interactive session. Demo Expo will last for 1.5 hrs in the Exhibition Hall. Each participant is allotted a table for the duration of the session. To participate, there is no need to have an accepted paper at the conference.”
Hello Robot is going to have multiple Stretches at IROS and we'd love to see some cool work from our community demonstrated at the Expo! We plan to make a few robots available for community members who would like to showcase their research at IROS, but aren't able to bring their own robots with them.
Demo submissions are due July 15 - if you'd be interested in using one of our robots to participate, please reach out to us as soon as you can!
Deep Reads:
Robotics Outlook 2030: How Intelligence and Mobility Will Shape the Future - Boston Consulting Group
This was actually published 2 years ago. Do you think the predictions stand the test of time?
A recap of the 2023 Cascadia Connect Robotics, Automation and AI Investment Conference in Pittsburgh - Cascadia Connect
Short snappy summaries from people at the forefront of robotics implementation. And saving the best for last!
AI for the Physical World: What's Next in Robotics, Perception and Foundation Models - the Innovation Endeavors x Google event last week.
We recorded the event and invite you to watch it here. In order of appearance, the speakers include:
Ryan Hickman (NVIDIA) - emergent capacities of NeRFs
Travis Deyle (Cobalt Robotics) - how LLMs will enhance our ability to do useful thing in human spaces
Johnny Lee (Google) - generative AI combined with AR, how human-in-the-loop can bootstrap new systems
Karol Hausman (Stanford) - with the advent of foundation models, what is the new ‘bitter lesson’ of robotics?
Frank Dellaert (Verdant) - for robust industrial use cases, what is newly possible and how do we run all of this emerging tech at the edge?
Jim Fan - (NVIDIA) - can we use video games as a means to crowdsource generalist machine capacities?
As usual, the Q&A may have been the best part! Please see a summary of some of the fascinating topics discussed below:
General purpose vs specialist robots
Today, we're able to build specialist robots that can get to high nineties performance using reinforcement learning and high-quality data from a specific domain. We know how to do this. However, the challenges for generalist agents - a robot that is good at many tasks - are significant, and we likely have not unlocked the correct techniques to achieve these capabilities.
Multi-task RL is very difficult, and multi-task imitation learning doesn’t get you to the performance you need. How we solve this is not yet clear.
Furthermore, in terms of applied capabilities, there are many examples where location, connectivity and operational complexity are the limiting factor - not the technology itself. Many outdoor use cases have engineering nuances that are the true blockers as opposed to research frameworks.
NeRFs
Standardization is key. For example, if we want to be able to use NeRFs for physics-based tasks, we have to decide on the rules. Today the research is split among material properties, dynamics, articulated, animated, rain & weather, and more. Hopefully we can settle on a standard so we can build toward a common goal. The temporal element of NeRFs is also a major gap.
Future methods and data
The programming of the future may be fine-tuning.
You’ll start with foundation models, but let’s say you want to run on lower power, or you want high percentile performance on your particular domain etc. — if you’re not preparing your infrastructure to fine-tune large models, you should start.
In a world with high-performance foundation models, most use cases will still need specialized data and capabilities, thus fine-tuning will remain a major opportunity. Ultimately, ask yourself - are you innovating on 1) the model, 2) bridging from digital to physical (and back), or on 3) unique data sets? All three will be needed in a world with amazing foundation models.
Data access will continue to be a challenge and a philosophical question.
How do you collect data from people's homes or private spaces such that you can make useful robots? LLMs are amazing with digital assets, but as long as models are starved of real-world data, it will be difficult to improve real-world performance.
Diffusion models tend to be slower than transformers, but can get you to a very precise action for spatial AIs. Nonetheless, the data collection challenges remain a blocker.
Language as the universal API, and other interesting ideas
LLMs just happen to be trained on language, but may not actually have much to do with language. Perhaps they are more of a universal computer, and pattern recognizer. Language has many benefits - it’s universal, expressive, and easily agreed upon - but it may not take us all the way.
Code is the highest class of language for foundation models, given its reasoning and controls advantages. Code is the way to map language onto the physical world. It’s also the way to make LLMs self-hosting, by letting a model autonomously rewrite its own code.
Taken to its end, language is transient. It’s too slow - it’s just a stop on the way to something much more efficient. LLMs, when asked to communicate most efficiently, do not respond with language but rather a gibberish that can only be understood by machines. Similarly, most animals do not rely on the organic equivalent of 4k cameras and speakers, there are many more compelling sensor modalities we may be able to explore.
If you’re working on anything in the space, want to connect, or have any questions, please reach out to Josh Rapperport at Innovation Endeavors