Machenta raises $40 million betting simulation can replace real-world robot training data
The robot training data industry is spending hundreds of millions collecting real-world footage of humans doing things. Mecka AI raised $60 million from Sequoia to record human motion. XDOF raised $70 million from Thrive Capital, a16z, and others to build data pipelines and annotation systems. We all know what Figure AI is doing. The field's dominant assumption is that building capable robots requires massive amounts of real-world physical data.
Machenta just raised $40 million betting that assumption is wrong.
After all, who needs pesky reality?
The London startup says its model is pretrained entirely in simulation and can learn a new task from a single real-world demonstration. No fresh training run. No data collection campaign. Show it once, and it goes.
The $40 million seed round was co-led by Creandum and Union Square Ventures. The 11-person team was founded by Imperial College London researchers Edward Johns and Vitalis Vosylius, and it builds on 6 years of Imperial research including work called Implicit Graph Alignment and Instant Policy.
What the VDA model actually does
Machenta calls its system a vision-demo-action model, or VDA.
(Just what we needed: more acronyms.)
Most robot training uses language instructions to specify tasks. The VDA takes a demonstration alongside live camera observations and predicts the robot's actions directly. The demonstration, basically, is the prompt.
Conventional task-specific training requires updating the model's weights using collected examples: you gather data, run a training job, wait. Machenta says its pretrained model interprets the demonstration directly, with no weight updates and no additional training run required.
That's the technical claim. There is an important asterisk.
The "instant" in Machenta's pitch refers to adapting to a new task, not to the substantial pretraining required to build the underlying model in the first place. That pretraining happens entirely in simulation, which is exactly the work the $40 million will fund: more compute, more simulated data, larger-scale training runs. The one-shot deployment is the output of that investment.
Similarly, Machenta's claim of using no real-world training data applies to pretraining only. The demonstration and camera feeds used at deployment time still come from the physical world.
What it can do right now
The company's current demonstrations include tidying tools, tending machines, and a 3-robot sequential assembly task. The same model runs across 6 different robot platforms, various camera arrangements, and multiple teaching methods: physically guiding an arm, using a VR controller, and combining shorter demonstrations into longer sequential tasks.
The reliability numbers Machenta reports are impressive for their scope: 74 block placements out of 74, 78 screwdriver handovers out of 78, 97 bottle-standing attempts out of 99. Note, these are company-reported results on a small set of self-resetting tasks, not third-party evaluations across a broad task library. That's normal for a research announcement.
The caveat is that those numbers don't tell you how the model behaves on tasks it hasn't been specifically demonstrated.
The company has also run early experiments guiding humanoid hands with human hand demonstrations. Co-founder Johns calls this preliminary and outside Machenta's main focus. They aren't claiming whole-body humanoid capability: manipulation on a stationary arm and manipulation by a mobile bipedal humanoid are very different engineering problems.
The bet against the data arms race
The dominant paradigm in physical AI right now is data collection at scale.
Physical Intelligence, the company that raised $400 million, built its foundation on varied real-world robot data. Mecka pays people to perform everyday tasks on camera. XDOF builds the infrastructure to annotate and pipe that data to frontier labs. The assumption underlying all of it: simulation isn't good enough, and real physical diversity is what makes models generalize.
Machenta's founders bet the opposite: that sufficiently varied simulated experience can teach a model to interpret demonstrations in general, without requiring simulations tailored to each eventual customer task. If a manufacturing company wants to add a new assembly step, they show the robot once instead of commissioning a data collection effort.
The practical question is whether this actually reduces the barrier to deployment.
If teaching a robot a new task becomes a short demonstration session rather than a multi-week data project, the economics of automation change for operations that are small, variable, or frequently changing. That's a large slice of the manufacturing world that industrial automation has never reached.
11 people, no commercialization plans yet
For now, this is still a research company, and there are no near-term plans to sell to factories. The $40 million goes toward hiring and larger-scale simulation training, and the team stays focused on proving the approach works across harder tasks and longer operational windows.
What's notable is who backed them: Union Square Ventures and Creandum don't typically fund research-stage companies that are many years from revenue unless the underlying thesis is genuinely compelling.
The idea that simulation-based pretraining can substitute for real-world data collection is a big bet, and it's not obviously wrong.
The real test will come when Machenta's model encounters tasks that were underrepresented in its simulation training, or environments that look different from anything it was trained on ... simulation-to-real transfer has been a persistent problem in robotics for a decade. What's changed, perhaps, is the scale and quality of physics simulation.
Whether that's enough to make Machenta's approach generalize is the $40 million question.
Sources
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”