Part 4: So I want to start a robotics company in 2024…

The recent explosion in AI has ushered venture dollars at full speed back into the robotics category, with many making the assumption that we’ll quickly see a similar breakthrough in the physical world. While we will forever be techno-optimists at Bison, we are also pragmatists, attempting to understand when the moment is right for such cataclysmic shifts. Today, I’ll dive deeper not into what the right use cases are for industrial robotics but into what we believe the most successful companies will look like.?

Earlier in this blog series, I wrote about the breakthrough that OpenAI had with its Dactyl model, which showed generalizable learning applied to dexterous movements with a humanoid hand. What I didn’t mention was that despite leading the way, OpenAI stopped investing in robotics shortly after showcasing that work. Sam Altman, CEO of OpenAI, explains that decision here in a conversation with Bill Gates. Sam muses, “We started robots too early, so we had to put that project on hold. It was hard for the wrong reasons; it wasn’t helping us make progress with the difficult parts of ML research… We realized more and more over time what we really first needed was intelligence and cognition, and then we could figure out how to adapt it to physicality. And it was easier to start with that.”

What Sam alludes to but doesn’t say directly when saying “it was hard for the wrong reasons” is that they found hardware limitations drastically slowed their rate of progress. The hard part was dealing with insufficient hardware (actuators, motors, etc), and they’ve now experienced better ROI from investing in teams directly moving hardware capabilities forward than from those developing them in house; their recent investments in 1x and Figure suggest they are sticking to this approach. This also means that they are confident they will be the ones to build the AI systems that will enable these companies, and much like Sam’s rumored fundraising for ChipFab, these investments by OpenAI should be viewed as an attempt to fund the ultimate consumers of OpenAI’s technology.?

Anyone that has spent any amount of time interacting with one of the AI chatbots or the thousands of derivative/wrapped products has experienced their fair share of the resulting agent hallucinations. When you are writing marketing copy, summarizing sales calls or drafting initial code, these edge cases of unreliability are tolerable — you just discard and reprompt (though I suspect we will see a number of high-profile flubs in the coming years, where people get a bit too lax with their fact-checking of chatbot-generated content).

When your model controls a thousand-pound device capable of causing massive amounts of destruction, unreliability becomes a significant issue. Robots in deployment require reliability levels of 99.9% to 99.999%, along with protocol for safe failure when the situation necessitates it. Not only are there significant hardware and data challenges to building generalizable robotics models and robotics platforms, but the bar is also significantly higher than it was in the case of the chatbots.? Our friend, Brad Porter, at Cobot further expands on this challenge in his blogpost on the topic of why humanoids may be farther away than we think and what a more reasonable path may be.

In the last wave of robotics investing, the consensus was that to build the most capable robot, you needed to own the entire stack soup-to-nuts: controls, perception/autonomy, hardware and software. This meant to build a robotics platform, you had to raise potentially hundreds of millions of dollars, and you still might find that there wasn’t enough. Today, the pace of change across these three swimlanes means the best companies will be those who most effectively decide, based on their use cases and business models, where it makes sense to be an integrator vs an innovator.

The rapid growth in robotics deployments and general market interest in autonomy also means that for the first time ever, a healthy enough horizontal ecosystem exists to enable robotics companies. If you were to start a SaaS company today, besides getting hundreds of cold emails from VC’s asking how they can be helpful, you’d also have the luxury of a healthy ecosystem of developer tools and products to choose from. We no longer expect the founding teams of SaaS companies to have expertise across the entire software stack. Identity management? There are many solutions for that — buy one. Payments? There is a whole industry for that — get a license. Data infrastructure? A non-engineer can now have a data pipeline set up in a single afternoon. The list goes on and on and, as a result, the founding team can make integration decisions rather than building tooling. As a result, software founders can focus the bulk of their efforts on the customer problem they are trying to solve, not on the piping to enable it. The same environment exists for e-commerce, web apps, mobile apps and most other healthy venture markets (the increasing virtualization of drug discovery is a good example of a similar ecosystem developing in a physical world industry).?

In the next five years, we expect the robotics landscape to undergo a similar transformation, allowing teams to be acutely focused on the core mission of their company, whether that involves addressing a specific industry problem, developing a class of robotics or refining a control system. It also means there will be, by necessity, a set of horizontal tools that will deliver these capabilities and enable the industry at scale.

Smart teams will integrate these tools, leverage the advancements in hardware and supply chains ushered in by adjacent industries and rely on externalized controls, perception and autonomy models. As a result, the robotics companies of the next decade will be 10x more capital efficient than those of the last decade. It also stands to reason that they will be 10x more time-efficient: Similarly to the way SaaS companies can have a demo-ready product within a sprint cycle of a couple weeks, robotics companies should increasingly be able to deliver MVP products to pilot customers on drastically reduced timescales.?

These companies will build better products and reach scale more quickly. For too long, teams have spent years maniacally focused on building a product (and overcoming significant technical hurdles) only to realize, upon initial deployment, that either the industry had completely moved or the problem wasn’t a significant enough pain point to begin with.?

This unlocked productivity and compressed feedback cycle will result in a virtuous cycle where robots get better, scale because they’ve gotten better and get better again because they’ve scaled.?

What does that mean? The pragmatist’s view on what a winning robotic company looks like in 2024 and an RFP

As you can tell by the prior posts in this series, we are bullish on the potential of robotics companies to have significant impacts on the physical world and drive profitable outcomes in the process. Yet, we remain deeply pragmatic about what the right phenotype of a company is to back. You can read more about our general theses in my colleague Tom’s blogs, “If you build it they WON’T come” and “10x or Bust.” In this post, I’ll briefly explore what we think the winning recipe is for a successful 1) applied robotics company 2) robotics infrastructure company.?

Applied Robotics Companies: These are companies that deliver a robotic solution to a customer. Whether their form factor is humanoid, wheeled, a drone or a fixed arm — or their business model involves device sales, Robotics as a Service (RaaS) or others — these companies ultimately achieve scale and generate significant revenue directly from their install base. To generate the scale of returns that get a venture investor like Bison excited, these companies ultimately need a path to hundreds of millions or potentially billions of dollars in annual revenue. It’s impossible to predict exactly what the right company will look like, but we believe it will have several knowable characteristics:

• A generalizable platform tackling a widespread, current pain point: Too often we see robotics companies building highly specialized hardware and software solutions to meet a critical need, only to realize during our diligence process that while they might be able to capture 100% of the market for that need, the market simply isn’t large enough. This is made worse when the incremental technical investment to unlock adjacent tasks isn’t incremental at all. This might look like a company tackling a single specialty crop and realize that the move from strawberries to blueberries requires a fundamentally new platform or a warehouse autonomy company finding that machine-tending use cases don’t make you any more prepared to solve material-handling use cases. The most exciting companies will not only start with use cases that are large enough to support hundreds of millions of dollars of revenue, but they will also build their platforms to be modular and flexible enough to expand their capabilities without a complete rebuild of software or hardware.
• A product competitive in labor costs (Bill of Materials and capabilities): We expect that to begin to ease labor shortages, robots must do undesirable tasks as effectively as humans can, and they must do so at a lower cost. This means that if your robot is half the speed of a human, it also needs to be less than half of the operating cost. In our estimation, this means that the total BOM for a robot competitive in labor costs can’t be more than $30-40k. In other words, a robot with a usable lifespan of five years and 80% uptime can achieve a 75% gross margin relative to a $20 per hour, single-shift employee over its lifetime.?
• A pathway to fielding robots within one year following investment: As discussed in our last blog, we believe there is a significant virtuous cycle that commences with initial deployments. This doesn’t mean it makes sense to deploy for the sake of deploying, but it does mean teams should be acutely focused on creating a Minimally Loveable Product (MLP). Once deployed, teams can iteratively adapt and add capabilities, akin to software products continually releasing features.??

Horizontal Robotics companies: These are companies that sell developer tools to other robotics companies. They build the shared infrastructure that makes it simpler, cheaper and quicker to achieve the above MLP. For these companies, growth is limited by the overall install base of their customers. However, we believe we are nearing a tipping point where the install base growth is significant enough to support an ecosystem of these companies over the next decade. So what will they look like?

• Emphasize a software-first approach: While there have been a number of recent monetization events for sensor companies (like the LiDAR SPAC craze of 2020-2022), sensor technology often leads to commodification in pricing. Consequently, we are more focused on software solutions that offer high gross margins and the potential for limitless scalability and innovation. This might mean owning a multi-modal foundational model focused on physical world interactions, or it might mean owning a much narrower slice of the stack.?
• Specializes on discrete problems common to all robotics companies:? It’s impractical for every new robotics company in robotics to have expertise in all aspects of autonomy. Smart teams will understand the nuance of determining what is critical enough for the applied robotics companies to own the codebase vs what can be abstracted. Examples of areas that might make sense to abstract include safety modules, path planning and sensor fusion.?

If you are considering starting or have already started a company in the autonomy space, we’d love to hear from you. If you look like the companies we describe above, we’d love to explore a partnership. If you disagree with us, we’re also open to having our minds changed.?

Please reach out: I’m at [email protected]