Rhoda

2026-03-13

tl;dr I believe Rhoda has the best shot of the robotic foundation model companies at solving generality and creating an enduring business. This comes after almost a year of researching the space, building projects, writing deep dives, and meeting folks across robotics. We are able to do this as a function of the way we’ve formulated the robotics problem and our relentless focus on commercial deployment. If you are interested in learning more about Rhoda, or want to work together (in any capacity), please reach out to me on X or via email.

My personal journey

My over-arching goal has always been to help deploy millions of robots into the world. I’ve seen robots not only help people physically, but also bring joy, inspire students, and build communities. I genuinely believe that once the “clankers” are here, mostly everyone will be happy they’re around. The main challenge with robots today is that they are very rigid, and can only solve overly-specified problems with precise definitions of success.

Mid-2025, I had the inkling that the most important problems in the world were being solved at the frontier of general purpose robotics. I had been tracking developments in robot learning since 2024 (check out some slightly cringe old blog posts) and was pretty confident that advances in robotics would soon be seen across the world. The rate at which models were improving was too fast for that not to be the case.

I knew that I needed to work at an organization focused on deploying general purpose models into the physical world, but my background in traditional computer vision and robotics (perception, planning, control) seemed to be less and less useful in a world where the Bitter Lesson had taken over. I wanted to work somewhere I could become the best at my craft, and where my work would endure beyond the venture cycle into production.

There were three big obstacles:

  1. I couldn’t seem to find a company that cared enough about diffusing robotics into the economy that also
  2. had a strong enough research backbone to not be eaten alive by existing foundation model players in 5 years and
  3. was willing to take a bet on me (a non-phd/postdoc in robot learning)

So, I quit my job and spent 8 months working on projects to build the skills in robot learning. This included understanding the bottlenecks (deep-dive, umi), reinforcement learning for physical robots (atari), and some tiny projects in computer vision (dft, diff-comp).

An Asymmetric Bet

My months of exploration meant that I had a chance to meet with folks across the robotics world. I learned about incredible companies building in HW/SW/AI - task specific robots that solve infrastructure challenges, data collection companies putting millions of workers and cameras out into the world, and even tiny labs working on new methods for self-play based reinforcement learning. I learned about approaches to actuators, camera/sensor selection, teleoperation, model evaluation, training paradigms, and ways to scale up compute and data.

In my opinion, there are N(~10) companies in the world that are best positioned (talent, resources) to solve the general purpose robot problem. Most of them, however, are attacking it along the same vector: (VLM -> VLA) + collect as much diverse robot data as possible. The data collection methods vary: on-embodiment vs off-embodiment vs sim vs world model. The product strategy varies: consumer vs industrial. The form factor varies: bi-manual vs humanoid. But under the hood, off-the-shelf VLMs like PaliGemma are the core stepping stone.

Rhoda is an asymmetric bet because our focus is on video generation as robot policies. If this ends up being the dominant way to train policies in the long run, then not only will we be best positioned to deploy, but also we’ll have shown that on-robot data is not the bottleneck in robotics.

Why Rhoda can win

  1. We have the most data-efficient robot policies that I have seen (per on-robot hour). Official blog post.
  2. We are well positioned to get robots into the economy. See our website for more!
  3. Strong mission-aligned research/eng team. Everyone at Rhoda is a joy to spend time with.

Robotics does not have an internet scale dataset built by 50+ years of accumulated human experience. Individual companies cannot realistically finance creating this. We take advantage of the best dataset humanity has ever created, the internet, and formulate it in a way that solves robotics. This has allowed us to build dextrous manipulation policies with less than 10 hours of on-robot data - check out this video.

Second, from a business perspective, our research objective is not bottlenecked by our commercial objective. Most general purpose robotics companies get stuck because each deployment saturates after N hours, meaning they can’t get extra value out of their main mechanism for achieving generality. In contrast, Rhoda will get stronger with deployments. Deployments help us build our operational muscle and bring in recurring capital so we can focus on mining the internet and training models.

Anyway, there’s a lot more I wish I could tell you about Rhoda, but an essay isn’t the right medium! If you want the full story, please come and work with us. Reach out to me on X or email and I can put you in touch with the right folks :)

← Back to Home