Robotics: ENACT raises nearly 27 million euros to develop the sense of touch

The Zurich startup is raising nearly 27 million euros in pre-seed funding to develop robots capable of better perceiving forces and adjusting their movements when in contact with objects. Its challenge consists of designing together the mechanics, the data and the artificial intelligence models necessary for manipulation. In France, Yann LeCun and Arthur Mensch are participating in the financing.

Handling a cable, using a tool, carrying out assembly: these actions appear in the Enact Intelligence data collection program. What they have in common is that they ask the robot to interact with matter, to accompany a movement and to adapt to what happens during the action. For the Swiss company, part of the answer lies in the information the machine can gather about its own movements and the forces it encounters.

Enact Intelligence thus intends to develop robotics capable of learning fine manipulation by combining AI models, sensors and actuators, the components that produce movement. The question relates as much to the design of the robot as to that of its learning: what information must be given to it so that it correctly adjusts its actions?

Learn what happens on contact

The startup describes its approach as a proprioceptive system combined with models capable of exploiting force information. Proprioception informs the robot about its own mechanical state; the perception of efforts should help him to adapt his behavior during manipulation.

The “sense of touch” here covers this ability to take into account physical interaction. Enact emphasizes forces and proprioception, without detailing its entire sensor architecture, so it would be premature to imagine artificial skin reproducing all the sensations of a human hand.

The objective is to enable the learning of manipulation skills that can be used on different tasks and in different environments. An ambition that we discovered in the analysis of the startup’s recruitment: future scientists must work on force-sensitive models, organize training data and evaluate behavior on real robots. Their program includes learning efficiency and robustness when moving to real-world conditions.

Mechanics enters the learning loop

Manuel Galliker, CEO and founder of the company, worked in particular in robotic control at 1X. He wants to rethink the data loop and how the robot body, through its actuators and sensors, can provide additional information to learn fine manipulation.

Enact is looking for skills in mechanical design, programmable electronics, real-time control and systems integration. The company also wants to develop the infrastructure necessary for training and evaluating its models.

This integration can make it possible to jointly develop learning materials and methods. It also commits the company on several fronts: a better model must work with reliable components, and mechanical modifications must remain compatible with the data and software used.

The breadth of the program explains the projects to which the financing will have to respond, thus the startup will have to finance AI experiments, manufacture and test equipment, then bring together the skills capable of making everything work.

Before autonomy, human gestures to record

One of the most revealing positions among those available is that of robot operator. Its mission is to guide the machines and remotely operate them using motion capture systems. These demonstrations must directly feed the training of the neural networks.

The work includes monitoring data quality, annotating trajectories, and reporting unusual behavior. The Enact program therefore provides for a structured activity for producing physical examples.

This organization places collection efficiency at the center of the project. How many demonstrations does it take to learn an operation? How much of this learning can be used elsewhere? How often should an operator take control? answers which will determine part of the cost of adapting the system to the needs of an industrialist.

In Zurich, several approaches to dexterity confront each other

Enact joins territory already occupied by mimic Robotics. The other Zurich company is developing an integrated approach combining robotic hands, learning models and devices for collecting human demonstrations.

Other companies are tackling the problem with models capable of controlling different robots. Physical Intelligence presented a general model trained on several platforms, while Flexion develops an autonomy layer aimed at different hardware. These approaches pose a common question: how much of intelligence can be transferred between machines, and how much depends on their physical design?

The choice of a manufacturer will also be based on its existing solutions: specialized automation, tailor-made integration or manual work.

The next milestone is measured over time

Enact is also preparing acceptance criteria and automated testing to support the move from prototypes to repeatable production.

For a customer, interest will be measured by pace, success rate, availability and necessary interventions.

The investors gathered cross several of these areas. redalpine positions itself at the intersection of software and science; firstminute capital particularly follows embodied AI and industrial technologies. Notion Capital includes in its thesis software that interacts with its physical environment. Their presence gives resources to the project; the next observable result will be an operation carried out sustainably for a client, with a measured operating cost. Enact Intelligence completes a pre-seed financing round of 26.7 million euros

According to Sifted, Yann LeCun and Arthur Mensch would also participate in the operation.