AI Meets Automation: Startup Visionaries on the Robotic Future
Robotics startups have begun gaining attention once again from investors through new advances in AI. The partnership Generative AI provides with robotics hardware could reshape the market for both founders and funders. As global population rates lower, especially in developed economies, the desire for robots to fill in the role of jobs that are labor-intensive, dangerous and time-consuming rises. New advancements are rising out of global markets, and here in the heart of Silicon Valley, Foothill Ventures is intimately investing in companies in the deep tech and software sectors.
Foothill Ventures hosted yet another fantastic event in Sunnyvale on Google Campus, this time a Robotics Founder & Funder Summit event that highlighted the trends, investments and future of AI in the robotics space. The startup panel was moderated by Sophia Yu from Foothill Ventures with guest speakers Max Cao, CEO and co-founder of Jacobi Robotics, Marin Tchakarov, CEO of Fox Robotics, and Ben Bolte of K-Scale Labs. Please see their bios at the end of the blog.
Here are our top takeaways from the panel:
- Data Collection Challenges: There is a heavy emphasis on the importance of data in improving robotic systems — having robots in the field to improve core products through real-world data collection is instrumental. Integrating a data flywheel into the engineering process is essential for feedback and performance enhancement, ensuring each robot contributes to the overall improvement of the fleet.
- Challenges in Hardware and Software Integration: There are no shortage of difficulties of integrating hardware and software in robotics. While hardware is essential, the key to success lies in overcoming the initial challenges of building and refining hardware to complement advanced software solutions. This iterative process is crucial for creating effective and reliable robotic systems.
- Founding and Fundraising Advice: Fundraising in the robotics sector requires a focus on building valuable products rather than catering to investors’ demands. When founders prioritize creating a useful, revenue-generating business, this naturally attracts investors. Maintaining focus, minimizing distractions, and following one’s passion are essential for successful fundraising and business growth.
Below is the full discussion from our startup panel.
Sophia: What approach do you take at your company, and why?
Max: Jacobi Robotics is a horizontal software platform that works across industries and across applications. The reason we chose to do this horizontally was because first, we looked at the hardware modality or the hardware form factor, which is robot arms. We saw that they have been hugely successful and currently, there’s about 4 million robot arms in the world right now working every single day. It’s a form factor that was very well explored, it’s very reliable and it can be produced at scale, so it’s getting a lot cheaper. Then we also saw that despite this sort of massive success in terms of deployment numbers, there is still so much potential for this form factor. There are so many potential applications where this can be deployed, but the barriers, the pain points, they now lay in the software. Deployment times are very long, reprogramming times are long, the machines aren’t networked, the software is archaic. So we founded this company to basically address that — to solve all of these software problems.
Marin: At Fox Robotics we integrate. We like to think of ourselves as a software company disguised as a hardware company. Any time there is a physical incarnation of the technology or of the software play in the real world, there is a chasm to be crossed. There are some miles of broken glass to be crawled through with that hardware product. Usually there isn’t a readily-available hardware product that just fits the use case, so one has to build it themselves. Try it out. Continue to optimize it, and refine the solution. Only then can one start to hand over pieces to the hardware component tree.
Sophia: Is there anything you do to solve the data flow problem? How do you ensure a great product?
Max: The question of data depends on which use cases you focus on. There are structured and semi-structured use cases, and then there are completely unstructured use cases. For the structured or semi-structured use cases, I think you can get pretty far with simulation. For example at Jacobi, our proprietary tech was this way to develop a very accurate and very general model of any robot arm, along with a very efficient representation of the robot workspace, which we can then embed into a deep reinforcement learning framework. Because of the optimality and reliability of the models, the agent converges very quickly and we get very high performance.
Then there’s the unstructured environment, and this is where I think a lot of the attention is these days — on foundation models for robots and general robots. And here, things get a lot more challenging because it’s no longer enough to just create a perfect model of your robot arm and a well defined workspace — you basically need to recreate the entire world in simulation. That’s very hard, very time consuming, and very expensive.
The other approach that we’ve seen a lot in research is teleoperation. That’s not exactly scalable and it’s very expensive to pay people to do this 24/7. Another challenge for teleoperation is thinking about flywheels. If you want to have a flywheel, you need to have positive value creation when you collect the data. The way to do that is you have a robot, do something useful, create some value, and then along with that, you collect some data. That way you create a scalable business model. There are some challenges there like finding the right use cases — I don’t think use cases in manufacturing are diverse enough to facilitate the type of data you want to collect. But that’s really the only way to get scalable data collection and sort of that flywheel.
Sophia: Is there a good way to have a fleet out in the real world collecting data?
Ben: I think you do need to have robots or whatever your product is, really in the field doing stuff to know how to get better. The question is how to capture the scale to improve core products. For example, Midjourney is a good example where Midjourney reached this large scale and then they have core images you can click on that maybe you want to upscale and they’re like, Oh that’s a good one, so you can use that. That’s a nice data flywheel. The problem is, like with robotics, it’s a very comprehensive problem. The data flywheel is very much integrated into the way you do engineering. So it’s good to be informed and cognizant ahead of time when you’re building the robot or whatever it is about what sort of data you want to collect and how each marginal robot is going to help feed back into the performance of every other robot.
Marin: At Fox we monitor the greatest performance through monitoring KPIs and performance in the field. We look out for any interventions that occur in the field anytime the algorithms return less-than-desired confidence levels. We have a whole data annotation team that basically reviews those logs after the fact. It looks for reasons why the performance is degraded. We take data collection very seriously, because ultimately, it impacts the performance of the product, the satisfaction of the customer and ROI. For us, that flywheel is very alive and well. It does create that kind of jet engine-like feedback loop. Performing in the real world we do get better and better over time and it’s critical to our operations as a whole.
Sophia: Are there any other challenges besides data that you think might be overlooked by technical founders that are very important?
Marin: Qualities like focus and tenacity are oftentimes spread too thin amongst brilliant teams out there. I’ve seen it time and again. Brilliant teams start solving a problem. They get to about 60% of the solution. It’s not even commercialized yet, or even anywhere close to being commercialized, and they get distracted with another shiny object because it’s right there. A lack of focus and therefore lack of execution in a confined timeframe is a killer. I would say to focus on that one institution until as quickly as humanly possible you are able to compress time and drive a solution to a market where people want to pay money for something that adds value. Then you optimize. Then you spread yourself over other ancillary or adjacent problems to solve.
Sophia: How is this new wave of AI affecting your product, your technology and your business? And what would your response be facing this?
Ben: I think the question is what’s the most natural way for getting AI to work in lots of facets of the economy? It’s kind of becoming more clear how to do that, which is what’s really changed in the last couple of years. I feel like working in robotics is kind of in your instinct, you want to take advantage of AI, what’s happening in AI because it’s happening and AI is transforming. Robotics is no different than one underlying approach, but it will be quite different in terms of how you translate that into the real world.
Max: With our background from Berkeley, we’re very close to the research. I do believe that long-term, this is the right way to go for robotics — one model to rule them all. But there are still challenges and I don’t quite see a clear path to getting there, where you can collect enough data, where the products can be reliable enough for a business model, use cases and all of these things…there’s still so many open questions. For Jacobi specifically, most of our deployments are in manufacturing logistics. For robot arm applications in those industries, one of the big bottlenecks that make it hard to develop new robotic arm applications is application logic, or the task planning. I think the short-term that can be sort of a quick win for this sort of approach were these large models, they could be sort of the task pattern to do the higher level reasoning and help these robots be adapted.
Sophia: As Fox Robotics was founded before the Chat-GPT moment, what are your thoughts surrounding adopting AI technologies into strategies in the future?
Marin: It doesn’t have any immediate aspect or impact on what we do. If we step back to plot and follow the curve of automation, saying x is the complexity of the use case and y is technological advancement, we’re going to see a relatively sloped curve progressing into the right. As we unlock better, faster, smarter technologies and stitch those together, we can obviously afford to solve much harder problems, disrupting every way we can. But by and large, it’s augmentative to what we do, it’s great because it lifts the awareness and interest in the space.
Sophia: What would be your tips and best practices in raising funds, as robotics is a pretty hard vertical to raise money?
Max: My opinion might be a little bit controversial, but it’s to not care about what the VCs want or are looking for. Instead, just build a good business that creates value, a good product creates value for the world, that is growing, has revenue, and is useful. And then the VCs will come up to you.
Marin: I would say being mindful about what it is that you do for raising capital to run your business. There’s this competition that unfolds in a space that tends to cause that distraction from not keeping an eye on the ball. At the end of the day, what founders need is money in the bank in order to run the business with the least amount of disruption. There’s a major value to time — it’s the most precious resource of all.
Ben: I think you should do something you actually want to do. I really like this quote from The Alchemist: “When you want something, all the universe conspires in helping you achieve it.”
Sophia: Is there any other advice you’d offer to founders or founders-to-be in robotics startups?
Marin: It’s very easy to say once there’s some success in one’s career, but it’s to follow your heart. Follow your passion. Be tenacious. Just being laser-focused on what you want to achieve.
Max: I think asking for help or getting help along the way in your journey. When you build something of value you’ll get investments and partner with other good investors, and they will actively try to help you find other founders, find others who you can ask for help. I found that startup ideas, they’re all very much not original at all. Other people have tried to do similar things in the past. There’s a lot to learn from other people’s mistakes, and what has helped me the most is asking for help.
Speaker bios:
Sophia Yu is an Investor at Foothill Ventures, an early-stage deep tech fund in the San Francisco Bay Area. She focuses on AI, enterprise SaaS, and robotics investments. Sophia holds an MBA from Yale School of Management and a Bachelor’s in Economics and Finance from Tsinghua University. For more, see her LinkedIn profile here: https://www.linkedin.com/in/pengyangyu/
Marin Tchakarov is the CEO & President at Fox Robotics. As a technology executive with over 29 years of experience, Marin leads Fox Robotics in developing intelligent automated solutions that drive productive, efficient, and safe warehouse operations. As CEO, Marin remains focused on expanding the company’s reach, drawing on his successful track record in growing operations and sales at leading tech firms including KPMG, Oracle, Jawbone, Pebble Tech and Kindred AI, among others. Prior to joining Fox Robotics, Marin was instrumental in FitBit’s acquisition of Pebble Tech, and he led the acquisition of Kindred AI, an innovator of AI-powered robots used for fulfillment, by global technology leader Ocado Group. For more, see his LinkedIn profile here: https://www.linkedin.com/in/marin-tchakarov-b914999/
Max Cao is the Cofounder & CEO at Jacobi Robotics. He holds a degree in Mechanical Engineering from UC Berkeley, where he was advised by Prof Ken Goldberg at Berkeley Artificial Intelligence Research Lab (BAIR), and Imperial College London. Max was a former consultant at McKinsey. For more, see his LinkedIn profile here: https://www.linkedin.com/in/max-cao/
Ben Bolte is the CEO of K-Scale Labs, a company building open-source humanoid robots. He has previously worked as an AI researcher at Facebook AI, where his work has been published in top academic conferences and journals, and as an engineer at Tesla AI, where, among other things, he wrote the CUDA kernels for the voxel occupancy foundation model used in both the autopilot and Optimus perception systems. For more, see his LinkedIn profile here: https://www.linkedin.com/in/benjaminbolte/
Acknowledgements:
A special thank you to our co-organizers and sponsors: CMU T&E, EchoHer, Google for Startups, Deel, GenAI Assembling, and the photographers and videographers: Vril Zhang, Yanhe Chen, Jackon Gong.
This blog is written and posted by Yuna Liang, Summer Intern 2024 at Foothill Ventures.
Follow us on LinkedIn: https://www.linkedin.com/company/foothillventures
Follow us on Twitter: @FoothillVenture
Foothill Ventures is a $250M technology-focused venture fund located in the Silicon Valley. We back technical founders across software, life sciences, and frontier technologies.
Questions, thoughts, reflections? Let us know in the comments below. We’re always looking for great entrepreneurs and early-stage ideas, and we’re always interested in having a discussion about venture, technology, and anything related. To see more about Foothill Ventures, please visit our website: foothill.ventures.
