Part 2: Why now? What’s different about the robotics landscape today?

This is part 2 of a series from Caleb Appleton on our investment team on why we at Bison Ventures are actively investing in and excited about the field of robotics and autonomy. You can read the intro to the series here.? If you are building in the space, we would love to hear from you. Reach out to Caleb directly at [email protected].?

As a team we’ve been looking at and investing in robotics companies for nearly a decade, including legacy investments in robotics specific businesses like Vicarious Surgical and other businesses such as Gingko Bioworks where robotics/automation was a key platform enabler.?

Still, invariably during our diligence of new robotics companies, the question is asked : is the company just too early or is the timing just right? We all know of a number of companies that were just too early – despite having a great insight at their core, they failed to take off because the technology, market or talent ecosystem was not yet mature enough. Friendster never became Facebook. Pets.com died in 24 months while Chewy did $10B in annual revenue in 2023. There were probably 30 attempts at grocery delivery before Instacart cracked it (and was turbocharged by a global pandemic).?

For a number of reasons, we believe that there has never been a better time to build a robotics company and that the timing today is right. There’s been a confluence of advancements across talent, technology and market willingness to adopt new technologies. Let’s dive into a few of these trends that give us confidence:

The Experts are here! While robotics has been a key area of intrigue for decades, the academic environment hadn’t yet matured to accommodate specialization in robotics. Students who wanted to be roboticists had to first become mechanical, electrical, or computer engineers and self-select into specialized training.? The earliest specialized robotics degrees came online in the late aughts and have rapidly accelerated over the past decade. This past year, Carnegie Mellon became one of the first to offer a dedicated undergraduate degree in robotics. We expect this trend to continue and accelerate. In 30 years, it will seem insane that there was a time when every leading engineering institution didn’t have a robotics program.

Deployments are finally happening at scale:

This supply of talent is likely being gobbled up by industry. Early adopters of robotics solutions are finally reaching scale – the global deployment of industrial robots has taken off in the last decade from just over 1M installed in 2012 to? just under 4M at the start of 2023.the teams at Google, Waymo, Amazon, Boston Dynamics, Cruise, etc. need top robotic talent to maintain competitive edges and deal with an increasingly daunting labor supply.

The hardware is getting cheaper (and better!): The vast majority of robotics tasks require speed, precision and strength.? For a robot to be commercially useful, it also needs to be able to operate with >95% uptime. For much of the field’s history building a device that could meet the demands of a commercial deployment necessitated exorbitant build costs that meant robotics were reserved for the highest of value use cases by the highest of value users (namely the government and giant companies). However, these tides are beginning to turn. While it is a safe assumption that new technologies will generally improve with time, several of the key components to building a commercial robotic platform have seen far greater than linear improvements in cost and capability in the past decade - driving down costs while drastically increasing capability:

• Batteries: When a robot has to be tethered to a constant power source (as most 6 DoF arms do today and virtually all humanoids will for the foreseeable future) the world of adressable tasks is limited to what can be done within a reasonable distance from a power source. As a result, much of the core work of a warehouse or manufacturing facility is off the table from the outset. The natural solution to such a challenge is running the robot off of batteries. However, lack of energy density and cycle degradation meant that this option was off the table until very recently (unless you wanted a robot that needed a battery swap every hour). Thankfully, the last fifteen years have seen a ~9x increase in the energy density of batteries that has coincided with an ~85% decrease in the cost of batteries. While there is still work to be done to get to a full-shift or multi-shift equivalent run-time, multi-hour runtimes and hot swappable batteries are opening the doors to new and increasingly valuable use cases.

• Actuators & Motors: Electric motors have become pervasive in our lives over the past half decade as they’ve been incorporated into every form of mobility from electric vehicles to scooters, skateboards and hoverboards. As a result, there now exists a global supply chain of both off the shelf and easily customizable components. This means that teams hoping to build a robotic system don’t have to start from scratch or buy the Rolls Royce of actuators as part of their build out and overall BOM costs should fall drastically.
• Sensors: The same is true as it relates to the sensor stack - any truly autonomous system needs a litany of sensors to drive perception and decision making. Onboard compute and sensing has dropped precipitously from a cost perspective. Just look at Lidar as a proxy, what cost >$50k just five years ago is now solidly 1/10th that cost. Similar advancements in capability with corresponding cost declines have been seen as a result of the mobile revolution driving down the cost of MEMs and image sensors/cameras.

Programming and controls: The biggest surprise is not that, as part of a VC blog in 2024, I’m going to talk about the impact of large language models and AI, but that it took me over 1,000 words to get to it. The past few years have seen drastic steps forward in how we control and program robots to do useful tasks. This has been driven by several key advancements:

• The Deep Learning Wave and Alexnet: When I first started making venture investments, we were one of the few funds focused on AI and Machine Learning and our focus, at that time was in leveraging the recent advances ushered in by the AlexNet paper in 2012 - suddenly any problem that required good image classification (diagnosis of medical images, autonomous driving, etc) seemed within reach. In the decade since, massive strides have been made both technically? (I now ride in a Waymo at least once a week) and from a usability perspective. It’s now possible for a quality team to have a basic autonomy platform up and running within a reasonable time frame utilizing open source models, training data and simulation environments. If you sufficiently constrain the problem to semi-structured environments and are willing to tolerate a high rate of disengagements a basic set of autonomous tasks can be completed in the matter of quarters. I recently spoke to a team in the heavy equipment space for large construction sites that was a shining example of this - they had deployed autonomous vehicles in less than 6 months by utilizing open source models, existing hardware and constraining the problem to a set of discrete and simple tasks within an existing protected environment. A decade ago, even with the same constraints, a strong team would have spent years to get to the same point.
• Reinforcement and Imitation Learning: Reinforcement learning, where you give a model a reward function and let it self optimize to complete a given task. This technique rapidly took off and for a period we saw a steady march of computers dethroning the best human competitors in Atari, Chess, Go, and DOTA. Before long, OpenAI bridged the gap to the physical world with their Dactyl work showing a robotic hand trained using reinforcement learning and a simulation environment proving that you could move quickly from a simulated reinforcement learning environment to completing tasks in the physical world. For the first time ever, we saw a generalizable programming solution with limited need for black and white heuristics and cartesian coordinate based rules.?Following on the early successes of reinforcement learning the field made further advancements via techniques like imitation learning (where you typically directly control the robot in a Simon Says/Follow the Leader type activity) and drastic improvements in simulation capabilities.
• Large Language Models: OpenAI’s GPT3, Google’s PaLM, Meta’s Llama transformed the world in 2022/2023 and the robotics field was not immune. While they aren’t, and won’t (yet) be a panacea for the world of robotics, we have seen early and clever manifestations of ‘RoboGPT’ to reduce the complexity of programming a robotic arm or platform. Figure posted an impressive demo just last week using an integrated version of OpenAI’s models to interact with a robot and the robot with its environment purely through speech commands. What used to take days/weeks of tinkering by a controls expert in a proprietary controls GUI can now be abstracted, at least partially, to natural language. We expect significant progress to continue to be made in this space in the coming years.

Taken together all of these advancements mean it is possible, for largely the first time, to build a robotic platform that is capable enough, cost competitive with traditional labor and attainable without having to raise 100s of millions of venture dollars and spend many years building prior to initial deployments.?

As we meet with the best teams building robotic systems, we are begging to see them embrace this exponential shift in cost and capability. So, why haven't we seen a drastic increase in the proliferation of robots? Well the answer is 1)we have and 2)it is still really complicated! We’ll explore what’s holding the field back and what we see as investable opportunities in the space in the next two blogs.?

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. So please reach out. I’m at [email protected]