Investing in: Robotics

Investing
Written By Matt Wichrowski

As I outlined in my 2024 Hard Problems post, my personal development experiment this year is to try out a more proactive, thesis-led approach. For my first deep dive, I chose Robotics, a perennially interesting but often disappointing segment of deep tech. The fascination with robotics has been a fixture of human imagination since the dawn of science fiction and occupies equal weight in utopian and dystopian narratives. But now, thanks to the breakthroughs in generative AI and language models, a renewed optimism has taken hold with the view that we might be on the cusp of realizing robotics’ full potential. This optimism has led many, myself included, to re-examine the segment with cautious excitement. This is my work-in-progress perspective.

False Starts and Lessons Learned

One could boil down past failures to the technology not being ready. While this holds some truth it overlooks nuance that should be considered to avoid the losses of false inflection points. If taken from the perspective of the customer there were three main limitations that created too much economic friction in the adoption flow: upfront investment, integration costs and downtime issues.

  • Upfront investment - Excluding the most scaled operations (auto / aero) no one could justify a six or seven figure purchase price. Balance sheets and margin levels were too weak to weather a major capital outlay, especially from unproven startups with inherent survival risk. Hardware has now broadly commoditized, resulting in far more cost accessible rigs. This cost compression plus the extreme VC preference of recurring revenue gave rise to the Robotics-as-a-Service (RaaS) business model which largely mitigates this friction by transforming robotics budgets from CapEx to OpEx. 

  • Integration costs - Setup time and environmental redesign was perhaps the biggest driver of robotic startup deaths over the past 10 years. The required weeks (or months) of on-site setup can destroy ROI estimates and tax the energy and goodwill of the buyer. Past robotics systems did not have the abstraction layers required to ensure turnkey solutions.

  • Downtime issues - Any Operations leader will know that downtime is 50x more expensive than any marginal efficiency gain. Apart from the most engineered environments, any introduced variability could bring a system to a halt. This could be anything from an abnormal part to a weak ethernet connection.

This time is different(?)

AI breakthroughs are now released on a weekly cadence with each unveil seemingly more sentient than the last. The advent of LLMs, transformer architecture, model chaining, etc. provide drastically greater powers of perception and semantic understanding. Robotic systems can simply do more and learn faster, creating the potential of true embodied intelligence within diverse and dynamic environments. This is either a paradigm shift or financial category creation or both. This is best represented with the release of Mobile Aloha, a fantastically impressive engineering feat that catalyzed a huge amount of pro-robotics fervor within my VC circles. While any paper will clarify that its research is relatively controlled, it does not take much to see the how these newfound capabilities directly address the limitations described earlier. Improved situational awareness and significantly expanded simulated training capacity have made systems far more resilient and adaptable to novel stimuli.

However not all inflection points are technical. The COVID pandemic showed the developed world just how woefully fragile international supply chains were with supply shortages and resulting price surges impacting everything from automobiles to baby formula. This led cultural focus and political power being redirected to local / national resiliency if not isolationism across the West.

The most significant geopolitical repositioning over the last five years has been China’s transition from economic rival to adversary. For 40 years Southeast Asia was the global workshop and served as the foundation for globalization. The CCP is now challenging the defacto authority of the US / EU with increasingly saber rattling on both sides. The AI race and seminconductor controls effectively serve as a capitalist proxy for the tightening diplomatic tension. Entire manufacturing strategies have been redesigned and financial markets have drastically withdrawn. Combined these factors have brought about an industrial reshoring that we’ve not seen in modern times.

Where is the opportunity?

I’m still very much in the discovery phase of this mental model but today I see two main buckets of potential: full stack vertical RaaS and “pure brain” OS.

Full Stack Vertical RaaS - This should come as little surprise to anyone who’s explored robotics before but the failures of the past have established a pretty clear roadmap for market entry. There are a myriad of areas that could be automated / enhanced but it’s the trick is finding uniquely asymmetric upside. As a result I felt it more prudent to establish a mental model for which applications make more or less investment sense, rather than a shortlist of predictions. Consideration factors include:

  • Environmental control - Despite the increase in adaptability we’re still learning what the limitations of Gen AI are. All things equal a more engineered / stable environment will result in faster training and lower downtime. The greatest success stories of large-scale automation have been in circumstances where the entire environment could be engineered around the robotic capability (e.g. auto production).

  • Safety risk - Failures / hallucinations will take place. If those setbacks result in safety concerns (human life, financial performance or environmental impact) it can result create significant adoption friction.

  • Throughput unlock - Any automation narrative will boast the massive cost savings but in reality this is only really true where there is a demand chokepoint, either in required skill or human capability. In my opinion this is the single most important factor within this shortlist and is the first filter I use for assessing opportunities. E-commerce pick and pack has been such a killer app over the last few years because it could quite easily unlock 24/7 operation.

  • Ease of Supervision - Despite recent advancements interventions will still be required so the proximity and convenience of “pulling in” a skilled human practitioner is important. You can almost think of this as an apprenticeship model, at least for a temporal amount of time during training. The more co-located a robot and human can be in their daily routine execution the less downtime and integration costs. It’s this reason that I feel some of the more obvious robotic applications have failed to take off - e.g. oil rig maintenance, subterranean mining exploration.

When applying these four “filters” the most exciting applications are wet lab automation (bio or chemical), intra-hospital logistics (e.g. blood samples, medication, supply closet management), defense and recycling. Fulfillment / warehouse automation remains interesting but the alpha has dissipated there. An alternative view is that the best bets are “reshoring leapfrogs” - industrial segments that are now being built from scratch in the West and can leverage best practice from day one (i.e. no replacement). Examples include semiconductor, carbon capture, battery production / recycling, etc.

Pure Brain OS - One could also argue that the market needs a sufficient level of infrastructure abstraction to properly enable the long tail of applications. Some use cases may not have the economics to support a full stack build but an underlying foundational model which can then be fine tuned for the last mile might. Much of the most groundbreaking research in recent years (see below) has been at this cerebral level with enhancing capabilities across perception, navigation and manipulation.

Intuitively this is the more attractive bet to make as a VC due to the distribution and cost advantages of software. It’s also the most obvious play with VC giants and corporate players already putting some massive cash to work. Like LLMs, if the required entry capital for this play is tens of hundreds of millions of dollars, there’s not really an option for smaller, pure play (pre)seed funds.

There are several open questions still live in my mind for this play:

  • Novel data acquisition strategies

  • Compute cluster requirements / existing hardware limitations

  • Pricing structures across inference

  • Local compute and batch-based transfer learning

  • Perception and navigation calibration across hardware

There could be an ecosystem / community / marketplace of models approach that usurps any standalone general purpose model. I personally struggle with this one, at least in the near term. Despite all the hype in robotics there’s just drastically less talent in the world right now that could contribute to said community. Then again HuggingFace was founded in 2016 when machine learning talent was a fraction of what it is today. I would think the most viable wedge for this could be either a reimagination of ROS or a modular architecture for robotics-specific model chaining. At the pre-seed stage this approach is uniquely dependent on a founding team with an exceptional network or community cultivation skills.

Additional Reading

Notable Research*

  • Mobile Aloha - mobile, bi-manual robotic training

  • No, to the Right - adapting to natural language corrections

  • RoboCat - self-improving robotic agent

  • GOAT - universal navigation system

  • RT-X project - open source X-Embodiment collaboration

  • SpatialVLM - vision-language models with spatial reasoning capabilities

  • PaLM-E - embodied multimodal language model

  • Genie - generative interactive environments

  • SIMA - AI agent for 3D virtual environments

  • RT-1 - robotics transformer

Notable Companies*

*In no way are these lists exhaustive. I just got tired of linking…

Matt Wichrowski
Next

Hard Problems I want to back in 2024