AlleyCorp Robotics | March Newsletter
AlleyCorp Robotics | March Newsletter
March 2024 | Report Card Season
Hey everyone,
Thanks again for tuning in to the newsletter this month! We’ve been anticipating that 2024 will be a ‘report card’ year as robotics companies head out to raise and investor squeamishness slowly recedes - thanks in large part to the AI boom (bubble?).
As March comes to a close, we want to use this forum to bring attention to funding highlights from the first quarter of 2024 and share examples of companies in the AlleyCorp portfolio that are finding success. After that, I’ll also share my best insights from the conferences I attended this month (MODEX and Nvidia GTC)!
Cheers, and reach out if you want to connect.
Brannon | LinkedIn | X | Substack
🏆 Updates from the AlleyCorp’s Portfolio
The Q1 uptick in funding has been a windfall for the most promising companies in the sector. We are excited to point to recent fundraising and commercial updates across our robotics portfolio (that have been publicly announced) in the first quarter of the year and illustrate the health of the industry.
Aescape
Aescape officially launched earlier this month. They are introducing the world's most advanced massage… powered by robots! 13 separate publications covered the launch, and users are loving it. Read more about it here.
Glacier
Glacier raised $7.7 million in new funding with support from a phenomenal group of investors including Amazon's Climate Pledge Fund, AlleyCorp (us! 🚀🤖) , New Enterprise Associates (NEA), Overture , Elemental Excelerator, and VSC Ventures. Read more about how Amazon made this investment as part of their ongoing commitment to support entrepreneurs and female-led companies in climate tech here.
Earth Force
Earth Force Technologies reached several significant milestones in the past month!
Commercially: They signed several new contracts, and are on path to meet their goal of tripling in 2024 from the 20 projects they deployed on in 2023
Funding: They are excited to bring onboard a new investor: American Family Ventures! The insurance industry has been rocked by years of climate-related natural disasters, including the types of mega-fires that Earth Force is focused on helping to eliminate.
For more on AmFam’s view of how technology can help reduce wildfire risk and loss, see their post The Vast Potential of Wildfire Tech.
Koop Insurance
Koop recently announced a pioneering Enterprise Risk Management (ERM) automation tool to enable security, simplify compliance, and save on insurance for their customers - Read more here! They are also Introducing their RISK EDUCATION SERIES - a series of blog posts that dive deep into various insurance coverages and practical tips for protecting Technology companies through risk reduction and transfer. Read the first post here.
Viam
Viam has raised $45M in Series B funding, helping further their mission to accelerate innovation across robotics, IoT, and smart devices. Union Square Ventures and Battery Ventures and others participated in the round. Read the full announcement about the raise here!
💰 Other Robotics Fundraises this Month
Biobot Surgical Raises $18M Series B: Biobot designs Mona Lisa as a robotic-assisted surgical positioning device for prostate disease management.
Bear Robotics Raises $60M Series C: Bear Robotics produces AI-powered indoor delivery robots that serve the U.S., South Korea, and Japanese markets.
Cambrian Robotics Raises $3.5M Seed+: Cambrian developed an artificial intelligence platform to enable robot arms “to surpass human capabilities in complex vision-based tasks across a variety of industries.”
RIOS Raises $13M Series B: RIOS develops and integrates AI-powered robotics for the manufacturing industry.
Automated Architecture (AUAR) Raises £2.6M Seed: Their mission is to build sustainable and affordable homes using its micro-factories.
Tales from my March travels
I was fortunate enough to visit MODEX and GTC over the last two weeks (alas, at the cost of missing SXSW, and Logimat which were reportedly very good). I have become further convinced that 2024 is going to be a monumental year for robotics. It feels like the approach to innovating in robotics has changed, and entrepreneurs are building more capable systems, faster, and with less money.
Public coverage for both conferences is ample, so instead I’ll share my key impressions and takeaways from each event and then point you to links of existing reports for further reading… 🗞️
📦 MODEX 2024
Image credits: MODEX2024 Website
MODEX 2024 did not disappoint! The convention center was packed with supply chain experts representing companies throughout the value chain. This year’s exhibitions were different from prior years because autonomous robots were front and center. Most exhibitors had some form of autonomous system actively operating while on display, whether they were an OEM, or end user (Locus robotics’s bots were unsurprisingly everywhere). Being at this show was a glimpse into the future of supply chain automation. More than anything, the takeaway is that automation of warehouse operations is truly top of mind for robot purchasers in the industry, and companies of every size are putting serious financial backing behind the goal.
This revelation came as a nice counter insight to the declining sales of industrial robots in 2023... in fact, it even supports it. Few companies at MODEX were showing off traditional, single point robotics automation solutions. Instead AMRs, smart robots, and other autonomy-first solutions were the darlings of the show. I was intrigued by the autonomous truck loading and unloading solutions in particular. These demonstrated the new way of building in robotics and many of them fit my definition of a “Productive Robot” or “Probot” - more insight on that here and a full report in the works.
My favorite trailer unloading Probots at MODEX
1. Slip Robotics
Slip robotics showcased trailer loading and unloading of palletized goods using their robotic truck bed. I tested out the robot and the experience was frictionless. I controlled the robot with a modified Nintendo Switch controller (and it was much easier to operate than any video game I’ve ever played). Familiar controls made it extremely simple to pan, or rotate (perfectly in place). When the robot was able to squarely identify the opening of the trailer, I had the option to enter an auto-homing mode, at which point the robot would autonomously make final adjustments and enter the trailer. Brilliant engineering - winning the show's “Best New Innovation” award.
2. Anyware Robotics
Anyware Robotics was another favorite of mine. They differentiated themselves from others in two ways. First, they are using nearly 100% off the shelf hardware, pushing the complexity to the software level. This enables them to rapidly iterate through the physical development of the entire system and opportunistically choose the most functional subsystems for their tech stack. When deployed in a warehouse, their bot can identify the trailer and get right to work without excessive oversight. Their second advantage is their proprietary conveyor system. With the mobile conveyor, the robot head does not need to travel to end of the trailer with every pick cycle. Instead, it simply pulls the boxes out onto the conveyor immediately in front of the boxes starting position. The addition of the conveyor more than halves the cycle time, and even more critically, boosts reliability since the robot does not have to carry boxes dynamically through free space (although of course it still can).
3. Pickle Robot Co
Pickle has been a player in the space for some time and it was impressive to see how far they have come even in the last 12 months. Their system was reliably unloading the truck throughout the demo, and they featured a live analytics dashboard to monitor system performance. They also are intentional about capturing the robotics controls data from every pick, enabling them to continuously improve their AI model.
4. Contoro Robotics
Contoro Robotics is a newer player in the space and is making rapid progress by leveraging teleoperations to improve reliability. The robot can of course autonomously pick up cases, recognize text, and palletize them, but for fast initial training and additional fault tolerance the operator can take control of the system remotely. Contoro is a great example of a lightweight and cost effective hardware build that is focused on solving the customer needs rather than building complex tech for the sake of it.
5. Boston Dynamics - Stretch
Boston Dynamics’s Stretch robot was almost cathartic to behold. It appeared to be the smoothest and most powerful of the non-palletized unloading solutions (witnessing BD’s engineering prowess is always inspiring). As an added bonus, Stretch routinely (and easily) unloaded 50lb+ boxes as part of the live demo, (which others were not showing). The consequence was that some of the heavy boxes were misplaced and fell off of the side of the conveyor before exiting the truck. Ultimately, at the of the demo a hired human operator had to go into the trailer to rescue the lost packages. Undoubtedly, they will use this data to improve robustness as time progresses!
Further coverage on MODEX 2024
💡🧠 NVIDIA GTC 2024
I made the local news talking about GTC - 📽️: ABC7 News Bay Area
This was the first time this conference was held in person since before the pandemic and the crowds showed up to show their appreciation. Nvidia (and AI in general) is having a very bright moment and they have no intention of slowing down the innovation. Nvidia GTC had something to interest every attendee - 3 giant exhibition halls, a full list of talks by the top minds in the field, developer training sessions, and of course - Jensen’s Keynote in the packed out 20k person SAP center (and filled overflow rooms back at the convention center 😯).
At the end of it all, I had two main takeaways (related to robotics):
Takeaway #1: Nvidia is dead serious about the robotic revolution
Jensen’s Keynote made ample mention of their intention to dominate the robotics space, where he delivered his famous line “Everything that moves in the future will be robotic.” Nvidia wants to build and become the base infrastructure layer that all advanced robotics are trained and built on top of (similarly to how they’ve leveraged Cuda to capture semiconductor development). The company is no stranger to taking big swings and this is another huge one.
You have likely seen that they announced Project GR00T (short for “Generalist Robot 00 Technology”) and their intention to create a robotics foundation model for humanoid robots. Product leaders at the company indicated that Nvidia’s intention is not for it to be competitive with other companies gunning to build similar tech, but rather to provide base competency for all robotics systems. As they describe it, they are performing “mission driven research”… but between you and I - I fully expect GR00T would naturally displace the (numerous) new startups chasing after robotics foundation models and if successful.
Takeaway #2: Artificial General Intelligence (AGI) for Robots is further away than we expected hoped
Jim Fan unveiled his initiative GEAR (short for “Generalist Embodied Agent Research”) and gave a phenomenally insightful talk on Generally Capable Agents in Open-Ended Worlds (this was my favorite talk of the show hands down). He highlighted the promise of Isaac Lab in the Omniverse and the role it will play leading up to the Foundation Agent. He built a simulator that featured an open world environment, and used a foundation model pre-trained on internet scale data. The results of which were very impressive.
We saw them use MineCLIP and Voyager to enable an autonomous robot agent in MineCraft begin to learn low level skills through exploration and take on more and more complex tasks in the game.
He then showed Metamorph project, where his team developed a single foundation model that can teach movement and locomotion to various body forms (thousands of different robot forms) and was able to adapt extremely varied kinematic characteristics.
Finally he showed us the new and improved Isaac Sim. Using this platform he ran many thousands of simulations to train agents to perform complex behaviors and to train advanced computer vision models. Through its use his team taught a virtual robot advanced martial arts (training 10 years in 3 days!) and trained another agent named “Eureka” simulated human-level dexterity (flipping a pen through its fingers continuously).
GR00T Foundation agent will be built by gathering robotic action data and associating it with task instructions for an embodied AI model (a robot). They will need to do this on a massive scale across 10s of thousands of simulated realities (somewhat similar to how ChatGPT was trained). The multi-modal model will learn from videos, language, and in person demonstrations, and teleoperated demos to develop skills in simulation as well as the real world. The grand vision of the project is that all agents, whether simulated or physically embodied, will take commands from the same foundation model and simply be different prompts. AGI wizardry 🧙🤯.
🤔 So why is AGI far off?
…Because there are still very large challenges to overcome.
There is no recipe yet for training robots. Nobody knows (or really agrees) how to properly develop skills across short term and long term learning systems (System 1 and System 2 type skills for the Kahneman initiated - R.I.P. to the legend).
Sim to real is a massive challenge that no one has figured out in robotics and will take time to compensate for. He did talk about some fascinating approaches they are taking to try to solve it, such as generating thousands of simulated realities with varying physical characterizations (tweaked gravity, modified friction, air density etc.) to build robustness into the model… but even Jim was uncertain if that can capture enough variation to work 🤷
Relatedly, scaling up data is almost impossible today. There is not a way to scrape robot control data from the internet (or physical world), and every data source we use to train them has strengths and weaknesses. Internet data for computer vision and VLM models was many orders of magnitudes more available than the kind of action data we need to make robots a reality.
Finally, even when you have a bunch of data, scaling up is still really hard. There is not a clear way to apply advanced models to a robotics stack even with all of the data in the world, or how to match those with robotics actions models well.
Fortunately, all of these problems feel naturally solvable with time. So I’m optimistic that foundation models for robotics will get there, but it's going to be a while before the robots do my laundry.
Further coverage on Nvidia GTC 2024
That's all we have for now. Don’t forget to subscribe and share - we will be back next month! In the meantime don’t hesitate to reach out via email, substack, LinkedIn, or X.
Keep building,
AlleyCorp Robotics team (Abe & Brannon)