Machenta 融资 40M 种子从一个演示中教机器人新任务

供资

Machenta 融资 40M 种子从一个演示中教机器人新任务

在 Google 上的首选来源中添加 Unite.AI

教机器人做新工作比自己干要长得多。 收集演示材料、编制培训数据并更新一个模型，可以将一个小的工作流程变化变成一个实质性的工程项目。 Machenta拥有4000万美元的种子圆形融资来缩短这一过程，开发出一个机器人地基模型来从单一的演示中获得一个物理任务而无需额外的特定任务训练。

伦敦宣布公司在10月27日的融资，由Creandum和Union Square Ventures共同领衔。 它将利用首都扩大其模式，并在国王十字的新办公室和实验室扩大团队。

Machenta称其方法为“即时机器人学习 ”。 称公司演示了包装、装配和清理工作方面的技术，机器人在被演示一次后适应变化。 雄心是广泛的，但证据应该被解读为一种正在发展的技术的演示，而不是证明机器人已经能够完成所有可能的任务。

没有新培训就学习新任务

关键区别在于培训一种模式和教授一种已经受过训练的模式。 Machenta仍在训练它的地基模型。 它旨在消除的是每次机器人遇到新工作时需要收集出一个大型的新数据集并运行另一个训练周期。

称教授、马申塔的共同创始人兼首席执行官Edward Johns在公告中表示

提议的工作流程刻意简单：展示所期望的行为，让机器人观察，然后让机器人执行任务。 在实践中，这将使向现有模式提供的指导成为示范的一部分，而不是单独的模型开发项目的开始。

当任务大体相同但对象或其安排发生变化时，这一点尤其重要。 只能重复固定序列的机器人需要环境符合其程序。 一个有用的学习者必须查明示威要求它完成哪些工作，并根据面前的局势调整行动。

模拟用于培训学习能力

Machenta的训练策略将大部分数据生成负担转移到模拟中去。 根据发布内容，目标不是列举机器人最终可能面临的每一项任务。 要使该模型暴露于足够多的不同情景，从而发展出从新实例中学习的可重复使用的能力。

这改变了公司预期规模会从何而来。 许多机器人学习工作严重依赖于通过远程操作或人类活动录音来收集物理演示。 Machenta打赌模拟经验可以提供学习新任务所需的多样性，在部署时提供教学的真正示范。

其技术介绍称系统为视觉-民主-行动模型，即VDA. 它需要演示和持续的相机观测作为输入，并以闭环的方式预测动作。 称公司其前期训练仅使用模拟；教学实例和活的观测是推论时间的真实世界输入。

区分问题：只进行模拟预训并不意味着部署的机器人在没有现实世界信息的情况下运行。 它仍然必须解释示威，观察其环境，并随着条件的变化选择行动。

内部机器人学习背后的研究

公司以约翰斯和联合创始人和伦敦帝国学院CTO Vitalis Vosylius的工作为基础。 约翰斯是帝国的机器人学习教授，两人从少数例子中发展出早期的学习身体行为的方法。

他们的"即时政策"研究在ICLR 2025上发表，描述从一两个演示中学习而无需再受训练的内文模仿。 它通过一个图表来表示演示、观察和行动，并利用一个经过学习的传播过程来生成行动预测所需的图表结构。

论文还描述了以伪演示来进行训练：模拟中产生的轨迹提供了一大批实例。 实验跨越了模拟和真实的任务，另外还有跨越机器人浮雕和语言定义指令的传输工作。

这项研究有助于解释科学基础，但并没有确定马申塔目前的商业模型具有完全相同的结构。 起步将较新的系统描述为从这一行工程所建立的基础模型。

内文学习从语言模型中是熟悉的，其中一个快取中的例子可以在不更新模型权重的情况下塑造反应。 将这一想法应用到机器人上的要求更高，因为一个输出必须成为现实世界所限制的物理动作序列，而不仅仅是一个可信的答案。

示威的建立—— 以及有待检验的

Machenta的技术员额报告持续了一小时的三项简单、自我重建任务：74次尝试中74次成功布局，78次螺丝刀移交，99次尝试中97次成功装瓶。 公司将两瓶失败归因于相撞风险和动能约束。

这些是有用的早期可靠性检查，而不是广泛的工业基准测试。 该帖子还介绍直接从人手学习和用人相接的手操作作为初步方向。 它没有规定每个工作场所的通用任务覆盖面或可靠的业绩。

对于未来的用户来说，令人信服的演示和可部署系统之间的区别包括可重复性、从失败中恢复、围绕人的安全行为以及机器人硬件的实际限制。 快速学习一个指令不能让一个抓取器处理一个它无法实际抓住的对象。

然而，商业机会是明确的。 如果教学时间大大缩短，就更容易考虑机器人的工作，因为这些工作变化太频繁，难以为冗长的编程和整合提供理由。 这既会影响工厂的部署，也会影响难以获得专家机器人支持的小型应用。

以种子回合衡量模型和团队规模

Machenta正在加入对适应性机器人智能的更广泛投资浪潮。 Unite.AI对梦幻实验室的跨机器人AI融资的报导，说明在不同的机器上，Machenta强调，新的物理教学能有多快成为有用的行为。

Creandum的搭档Kathi Wilhelm在组建公司前指出了团队多年的研究。 Union Square Ventures的管理伙伴Rebecca Kaden强调机器人的价值，这些机器人可以迅速接起新的任务，而不是要求为每项工作单独培训。

对于Machenta来说，下一阶段是将研究基础转变为更广泛、更可靠的能力。 这4000万美元给了公司资源来推广模型并围绕模型进行招募，而其创始团队带来则在6年时间里为一发机器人学习入行。

中心的问题是，随着任务的变化和要求的提高，学习行为的规模有多大。 如果这个方法能坚持下去，那么重要的变化将是实际的：人们可以花更少的时间来建造一个培训管道，而更多的时间来向机器人展示他们实际需要完成的工作。

本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Funding

Machenta Raises $40M Seed to Teach Robots New Tasks From One Demonstration

Add Unite.AI to your preferred sources on Google

Teaching a robot a new job can take far longer than performing the job itself. Collecting demonstrations, preparing training data and updating a model can turn a small workflow change into a substantial engineering project. Machenta has raised a $40 million seed round to shorten that process, developing a robot foundation model designed to acquire a physical task from a single demonstration without additional task-specific training.

The London company announced the financing on October 27, with Creandum and Union Square Ventures co-leading the round. It will use the capital to scale its model and expand the team at its new office and laboratory in King’s Cross.

Machenta calls its approach “instant robot learning.” The company says it has demonstrated the technology on packaging, assembly and tidying tasks, with robots adapting to variations after being shown what to do once. The ambition is broad, but the evidence should be read as demonstrations of a developing technology rather than proof that a robot can already perform every possible task.

Learning a New Task Without a New Training Run

The key distinction is between training a model and teaching an already-trained model a particular task. Machenta still trains its foundation model. What it aims to eliminate is the need to collect a large new dataset and run another training cycle each time the robot encounters a new job.

“Intelligence is not so much about what you can already do, it’s about how quickly you can learn to do something new,” said Professor Edward Johns, Machenta’s co-founder and CEO, in the announcement.

The proposed workflow is deliberately simple: demonstrate the desired behavior, let the robot observe it, and then have it execute the task. In practice, that would make the demonstration part of the instruction given to an existing model, rather than the beginning of a separate model-development project.

This is particularly relevant when the task stays broadly similar but the objects or their arrangement change. A robot that can only repeat a fixed sequence needs the environment to conform to its program. A useful learner must identify what the demonstration is asking it to accomplish and adjust its actions to the situation in front of it.

Simulation Is Used to Train the Ability to Learn

Machenta’s training strategy moves much of the data-generation burden into simulation. According to the release, the goal is not to enumerate every task the robot might eventually face. It is to expose the model to enough varied scenarios that it develops a reusable ability to learn from a new example.

That changes where the company expects scale to come from. Many robot-learning efforts depend heavily on collecting physical demonstrations through teleoperation or human activity recordings. Machenta is betting that simulated experience can supply the diversity required for learning new tasks, with a real demonstration providing the instruction at deployment.

Its technical introduction calls the system a vision-demo-action model, or VDA. It takes a demonstration and ongoing camera observations as inputs and predicts actions in a closed loop. The company says its pretraining uses simulation alone; the teaching example and live observations are real-world inputs at inference time.

The distinction matters: simulation-only pretraining does not mean the deployed robot operates without real-world information. It must still interpret the demonstration, observe its environment and choose actions as conditions change.

The Research Behind In-Context Robot Learning

The company builds on work by Johns and co-founder and CTO Vitalis Vosylius at Imperial College London. Johns is Professor of Robot Learning at Imperial, and the pair developed earlier methods for learning physical behavior from a small number of examples.

Their Instant Policy research, published at ICLR 2025, describes in-context imitation learning from one or two demonstrations without further training. It represents demonstrations, observations and actions through a graph and uses a learned diffusion process to generate the graph structure needed for action prediction.

The paper also describes training with pseudo-demonstrations: trajectories generated in simulation that provide a large pool of examples. Experiments span simulated and real tasks, with additional work on transfer across robot embodiments and language-defined instructions.

That research helps explain the scientific foundation without establishing that Machenta’s current commercial model has exactly the same architecture. The startup describes the newer system as a foundation model built from this line of work.

In-context learning is familiar from language models, where examples in a prompt can shape the response without updating model weights. Applying that idea to robotics is more demanding because an output must become a sequence of physical actions under real-world constraints, not simply a plausible answer.

What the Demonstrations Establish—and What Remains to Be Tested

Machenta’s technical post reports hour-long trials 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. The company attributes the two bottle failures to collision risk and kinematic constraints.

Those are useful early reliability checks, not a broad industrial benchmark. The post also presents learning directly from human hands and operation with humanoid hands as preliminary directions. It does not establish universal task coverage or dependable performance across every workplace.

For a prospective user, the difference between a convincing demonstration and a deployable system includes repeatability, recovery from failure, safe behavior around people and the practical limits of the robot’s hardware. Learning an instruction quickly cannot make a gripper handle an object it cannot physically grasp.

The commercial opportunity is nevertheless clear. If teaching time falls substantially, it becomes easier to consider robots for work that changes too frequently to justify lengthy programming and integration. That could affect both factory deployment and smaller-scale applications where specialist robotics support is difficult to obtain.

A Seed Round to Scale the Model and the Team

Machenta is joining a wider wave of investment in adaptable robot intelligence. Unite.AI’s coverage of Dream Labs’ cross-robot AI funding illustrates another effort to make learned capabilities useful across different machines. Machenta’s emphasis is on how quickly a new physical instruction can become usable behavior.

Creandum partner Kathi Wilhelm pointed to the team’s years of research before forming the company. Union Square Ventures managing partner Rebecca Kaden emphasized the value of robots that can pick up a new task quickly instead of requiring a separate training effort for each job.

For Machenta, the next stage is to turn that research foundation into broader, more reliable capability. The $40 million gives the company resources to scale the model and recruit around it, while its founding team brings six years of work on one-shot robot learning into the business.

The central question is how far that learning behavior scales as tasks become more varied and demanding. If the approach holds up, the important change will be practical: people could spend less time constructing a training pipeline and more time showing a robot the work they actually need done.

Auto-fetched and translated by AI Gem Box. All rights belong to the original author.