London startup Machenta published a technical introduction on October 27 describing a model that learns a new task from a single demonstration without fine-tuning. The model is pretrained entirely in simulation. Alongside it, the company raised $40 million in a seed round.
Key takeaways
- The architecture, called vision-demo-action (VDA), takes a demonstration instead of a language instruction
- Pretraining happens entirely in simulation, with no real-world training data
- The same model handled six different robots and several teaching methods, including a VR controller
- Hour-long run results: 74 of 74, 78 of 78 and 97 of 99 successful attempts across three simple tasks
- The team numbers 11 people, remains in research and has no near-term commercialization plans
What is vision-demo-action?
A typical robot control model in the VLA family takes an image plus a natural-language instruction and predicts actions. Machenta swaps the middle term: instead of a text instruction the model receives a demonstration of the task alongside live camera observations. Hence vision-demo-action, or VDA.
The difference surfaces only once someone teaches a new task. Conventionally that means collecting a dataset of examples, running training and changing the model's weights. Machenta says its pretrained model interprets the demonstration directly, with no weight updates and no waiting for another training pass.
Mechanically this is close to what language models call in-context learning — new information arrives as input rather than as a change in parameters.
The diagram splits the two paths after the same demonstration. The upper branch is the cost VDA aims to remove: data collection and a separate training run.
What exactly does “instant” mean?
The word refers to adapting to a new task, not to eliminating training altogether. Building the underlying model still requires substantial pretraining — what changes is what happens afterwards, when deploying the next operation.
The claim of using no real-world training data should be read the same way. It applies to pretraining alone, making this a pure sim-to-real transfer. The demonstration and camera images used at runtime come from the real world.
One demonstration is also not always the end of it. The company itself describes adding examples when a task is ambiguous, stitching shorter demonstrations into longer sequences, and trying alternative strategies when the robot's physical limits block the first approach. That still falls within one-shot imitation learning, but with caveats.
What was actually shown?
The launch material covers manipulation demonstrations across six robots, with different camera arrangements and several teaching methods — from physically guiding an arm to using a VR controller. The examples include tidying tools, tending machines and a three-robot assembly sequence.
The company also reports hour-long runs on three simple, self-resetting tasks: 74 successful block placements out of 74 attempts, 78 screwdriver handovers out of 78 and 97 successful bottle-standing attempts out of 99. These are company-reported figures on a narrow task set, not evidence of equivalent reliability across the full demonstration library.
| Task | Successful attempts | Total attempts |
|---|---|---|
| Block placement | 74 | 74 |
| Screwdriver handover | 78 | 78 |
| Bottle standing | 97 | 99 |
Shown separately are preliminary?Preliminary experiment: A result obtained outside the main line of work, without full validation. It shows something worked once, not that it works repeatably. experiments in which human hand demonstrations guide humanoid hands. Machenta explicitly calls these preliminary and outside its main focus — they do not establish general-purpose, whole-body humanoid capability. The humanoid sequence video is labelled autonomous in the source and played at 4× speed.
Why not simply fine-tune the model?
A simpler route suggests itself: record a few dozen demonstrations and run fine-tuning. That is how most deployments work today.
The problem is economic, not technical. Every new task becomes a separate project: collect data, run training, validate, deploy. At high volumes that pays off, because the cost spreads across thousands of cycles.
At short runs and frequently changing operations the maths stops working. The robot sits idle because preparing it for a new job takes longer than the job itself. That is why automation skips small plants and stations retooled every week.
If teaching a new operation collapses into a short demonstration, those tasks come back into play. That is the real stake behind the VDA architecture, rather than the score on three test tasks.
Is it live yet?
This is a research result, not a product. Machenta has 11 people and — according to Business Insider — remains focused on research, with no near-term commercialization plans. The funding is to support hiring and larger-scale training: compute and simulated data.
The seed round came to $40 million, co-led by Creandum and Union Square Ventures. The company was founded by Imperial College London researchers Edward Johns and Vitalis Vosylius, and the approach builds on six years of work at that institution, including Implicit Graph Alignment and Instant Policy.
Broader reliability and commercial deployment remain to be established. The published figures cover simple, self-resetting tasks — the kind that can repeat in a loop without human involvement.
Why does this matter?
Machenta's bet is about something other than data scale. The company wagers that sufficiently varied simulated experience teaches a general ability to read demonstrations, rather than requiring simulations tailored to each customer task. If that holds at harder tasks, what changes is not robot performance but the cost of entering automation at all.
What's next?
- The funding is earmarked for hiring and larger-scale training — more compute and more simulated data
- The company states it has no near-term commercialization plans and remains in a research phase
- The next test is whether the flexibility survives harder tasks and sustained operation beyond research demonstrations
Sources
- Machenta — Instant Robot Learning
- Humanoids Daily — Machenta Raises $40 Million to Teach Robots New Tasks From One Demonstration
