Ten questions about robotics
From horizontal bets to sim2real to Nvidia—where is robotics headed?
Over the last year, we’ve spent a lot of time thinking about robotics in its broadest sense, from traditional robotics, to drones, nanobots, AVs, and beyond. There are several macro reasons for our interest, which have been pretty well documented elsewhere (see here from Eclipse, here from Felix at Playfair, and here from F-Prime), so we won’t repeat similar points around macro trends supporting robotics adoption; impressive breakthroughs in models like RT2/Auto-RT, Aloha, Pi; and the sophistication and cost reductions of hardware.
But in looking seriously at dozens of robotics startups and investing in several, there are a number of questions that we’ve been thinking about, which we thought were worth sharing. I’d love your feedback, especially if you have a different view!
Do robotic foundation models make sense?
The dream of horizontal robotics—where general-purpose robotics models can adapt to diverse tasks—remains a lofty goal. Hundreds of millions of venture dollars are being thrown at it, which given the prize is a model able to replace most of manual human labour, is perhaps unsurprising. But this is really a data race, and whether it’s possible to collect enough data to train models to a sufficiently robust level. Teams are leveraging teleoperation, imitation learning, kinesthetics, simulation and other techniques in the hope they can get there first and scaling laws hold. We don’t know if this will work, or whether the models will end up being lightweight enough for inference to run at the edge on-robot, though teams at Deepmind, Pi, Skild, World Labs and others are going to spend vast sums on compute to find out (and it’s great that Pi have open sourced their first model here). But there are other approaches worth watching too, like Opteran, who are building a neuromorphic foundation model that doesn’t require the same data to train by leveraging insights from how insects process information. And there are other possible ways to get creative about how to use data to train models given the lack of 3D data e.g. projecting a multi camera scene to a point cloud, predicting updates based on physics, rendering the prediction with NeRF, and training end-to-end with video prediction loss.
How should we think about vertical robotic solutions?
While horizontal bets around world models are capturing our imagination, narrower solutions focussed on solving discrete challenges are quietly making progress. We’ve built conviction in startups going after very large legacy markets that are desperate to automate, but where the technology has until now not been robust enough. A combination of factors including sophisticated and cheap hardware eg powerful edge compute, impressive sensor fusion techniques, bespoke high quality models, and quick and cheap installation, together mean automation is not just possible but inevitable in a vast number of industries—from material handling, to lab automation, to construction. We’re investing in each of those domains, and hope to see many more companies exploring other markets.
How important are LLMs for robotics?
The addition of LLMs into robotic architecture seems to have had a material impact. But while we think they could become powerful tools for chain-of-thought tasks, we’re unconvinced they are a panacea (yet!) for reasoning for more complex manipulation, or necessarily the right representation space in which to plan (vs something that explicitly incorporates concepts such as causality at appropriate levels of abstraction).
How far can the sim-to-real gap be closed?
We’re excited about the progress being made, but conscious of how much is left to do. Various approaches like domain randomisation, domain adaptation and system identification are driving progress. To give a sense of it: VSIM is leveraging GPU-based parallelisation for massively improved physics-based simulations, Transic is a novel approach using human intervention and online correction to augment simulation policies, LucidSim combines physics simulation with generative AI models and CycleGAN synthesises realistic depth images from simulated ones to help bridge the gap. And there may be specific tasks, hardware and architectures where the sim2real gap is smaller e.g. by leveraging depth data. Even if there is, we’re still likely to need more sophisticated physical models, improvements in the fidelity of simulated environments, and more robust transfer learning techniques that can generalise across a wider range of real-world scenarios. There’s a long way to go.
Does form factor matter?
The debate over whether humanoids are the most sensible form factor to build for rolls on. It clearly does make sense on one level, given the world has been designed for humans and when the world models arrive, plugging them into a humanoid would be profoundly powerful. But the reality is that we’re still a long way from humanoids having the required dexterity to really replicate humans, and the hardware in both sensing (robots lack touch and feel) and movement (still very mechanistic), are relatively basic and brittle for more complex tasks in the real world. This will no doubt improve over time. But as Jenny Read from ARIA questioned (see below), we’re also not sure why we’re limited to human form either; it seems to lack some imagination. So for now, it leads us back to companies building the form factor that makes most sense for the vertical they are solving for—this could still be humanoid, but the chances are it’s not.
What do customers care about?
There has been a lot of noise (understandably) about the breakthroughs in robotics models and humanoid demos that look great on X. But speaking to dozens of customers of robotic solutions, there’s a pretty clear hierarchy of demands: robustness, speed and cost. Much is made of 99.9% reliability in robotic solutions, but for a task that involves moving 2000 objects per hour, that still implies a good chance that something will go wrong multiple times a day and require human involvement. Human involvement means overhead, reducing the ROI, which needs to be offset by speed and cost for the maths to work. The good news for robotics startups is that we’ve been surprised on the upside about how much more sophisticated buyers are about ROI—it’s no longer just about labour replacement costs, but the impact on brand if orders aren’t fulfilled, the costs of training and re-training in industries where there are high turnover rates, and insurance. Proving ROI is still critical and it derives from robustness, speed and cost, but customers are increasingly willing to believe the case.
Do you need to be an end-to-end solution provider?
Robotic hardware is increasingly commoditised—a UR10 costs £20-30k and RGB-D camera costs £hundreds. By our calculations, around 75% of the total cost of a robotic solution goes to the systems integrator, and around 25% to the software and hardware providers. For startups, selling a software-only solution that fits into another company's stack, who themselves are only going to make 25% of the total cost, inevitably means getting squeezed. It also means being reliant on your customer to win contracts, and losing the ability to showcase the importance of the software layer in question. As a result, for early stage startups in particular, we see considerable value in providing the full end-to-end solution to final customers.
Sense, think, act – what matters most?
The general consensus is that we have good, cheap sensors and actuators, and startup value is largely accruing into the thinking (AI + software) layer. We agree with this today. But we also think there are limitations to hardware too e.g. the expense and function limitations of Lidar, the lack of dexterity in end effect actuators, etc. So we’re also very interested in novel sensing (e.g. using MEMs or bio sensing), and research initiatives like that being led by ARIA for smarter robotic bodies. As the thinking layer matures further, our attention will inevitably return to hardware limitations.
Unbundling or rebundling of the stack?
It used to be the case that the stack was owned top to bottom by the big players like ABB. That is no longer the case—you can buy sensors, actuators, microcontrollers off the the shelf, use open source tools for operating systems like ROS or HuggingFace’s LeRobot ML libraries, buy software packages for navigation/localisation, visualisation, simulation, fleet management and workflow. This is democratising access to building robotic startups. The challenge for startups (linked to a previous point) is whether a point solution in the stack can be sufficiently differentiated to capture value, or be applied to unlock a specific vertical use case.
The elephant in the room is Nvidia, which is not so quietly trying to rebundle the robotics stack around its various offerings. Jenson’s keynote at CES 2025 highlighted the importance they place on robotics and it is becoming a very powerful stack from data collection (GR00T-Teleop, GR00T-Mimic for demonstration learning and GR00T-Gen to produce variations), world models via COSMOS, training via DGX, simulation via Issac Sim/Omniverse, and deployment through Jetson/AGX and orchestration through OSMO. Those companies that can leverage this powerful stack and augment it to deliver high performance outcomes may end up having the best chance of success.
Are there any novel business models that make sense?
Our core belief that selling an end-to-end solution is table stakes for early stage robotics companies. But there have been some experiments around business models that are worth noting. Robotics-as-a-Service (RaaS) has struggled to gain traction—while the model promises flexibility and cost-efficiency for customers by turning capital expenses into manageable operating expenses, it often demands high upfront investment and continuous maintenance from startups to keep fleets operational and cutting-edge. We’re more excited about companies taking more novel approaches e.g. Electric Sheep has rolled up traditional landscaping businesses in the US and is integrating its grass cutting robots to improve margins. We expect to see more of these sorts of innovations, especially in verticalized solutions.
If you’re building in robotics, I'd love to hear from you!