Enact Intelligence is not another robotics startup betting everything on computer vision. The Swiss company emerged from stealth today with a $30 million pre-seed round and a different technical direction: making robots perceive force and touch through their own actuators, reducing the reliance on cameras and visual streams. The round is led by Redalpine, Notion Capital and Firstminute Capital, with angel investors including Arthur Mensch, the head of Mistral, Yann LeCun, Meta's former chief AI scientist, and Thomas Wolf of Hugging Face.

CEO Manuel Yves Gallier, who previously worked at Physical Intelligence, describes an approach in which robots "feel" what is happening in real time and learn more data-efficiently while improving reliability and performance. Enact is rethinking the full stack: model architectures for multi-frequency and multi-scale intelligence, data pipelines, and, at the bottom, the actuators and sensors that feed the model.

The stakes go beyond the funding, although a $30 million pre-seed is high by industry standards. In a robot that must grasp an object without breaking it, the perception-action loop closes within very tight time windows. If the force and touch signal originates in the actuator, it cannot travel to a remote data center and back: latency would make stable control impossible. That shifts the center of gravity of inference toward the edge, inside the robot or the industrial cell. This is not an entirely new dynamic for people working with local controllers, but it carries a different weight when control is handled by neural networks trained on physical signals. You need models light enough to run on embedded hardware, enough memory to manage multiple time scales, and a runtime that does not depend on a cloud API. For those evaluating on-premise deployment, the trade-offs are familiar: fewer round-trips, more predictability, more control over process data. AI-RADAR provides analysis and frameworks for these decisions on /llm-onpremise.

There is a second element that sets Enact's path apart. Cameras produce enormous numbers of pixels, but much of that signal is redundant for a grasping task. Force and touch are physical signals with higher information density for the job: they capture micro-slips, pressure variations, vibrations. A model trained on this data could in principle achieve similar performance with fewer examples and, as a result, with a more contained computing load during training. The emphasis on data efficiency is not a detail: it is a lever for making manipulation data collection sustainable, since that data is expensive and not always replicable in a lab.

It is no accident that the round includes names such as Mensch and LeCun. On one side, the world of large language models is looking for deployment surfaces beyond text and code; on the other, robotics needs more flexible representations of the physical world. Enact does not explicitly say it wants to build an LLM for robots. But the reference to multi-frequency and multi-scale architectures touches a structural problem: merging fast signals such as contact with slow signals such as planning is not a simple extension of language models. A runtime designed to generate tokens is not automatically suited to controlling an actuator within milliseconds. The success metric here is not throughput, but grasp stability and the ability not to damage an object.

For Enact, the funding will be used to expand the team and accelerate hardware and R&D. The bet is that a tactile, data-efficient path can move general-purpose robots out of labs, into factories and into the real world. That requires not only better models, but an entire platform where software and actuators are designed together. It is a signal that the next front of applied AI will be fought not only in data centers, but also in precision mechanics, sensors, and the milliseconds between action and reaction.