How it works

Robot learning in short.

Data sources feed one learning pipeline that serves every form factor. From the first demonstration to live operation.

Data sources

Engineer collecting demonstrations through teleop rigs on two robot arms

Teleoperation

Proven today

People demonstrate the task through teleop rigs, typically 50 to 200 demos per task.

Humanoid robot training a grasping task in simulation

Simulation and world models

Scaling now

Synthetic data multiplies the demos, roughly by a factor of 20, plus pre-training in simulation.

Workers wearing head cameras recording their hand movements in production

Learning from video

Still early

Robots learn from videos of people doing the work. Promising, but early.

Robot learning pipeline

  1. 1

    Collect

    teleop demos on your parts

  2. 2

    Curate & Augment

    clean, label, scale in simulation

  3. 3

    Train Policy

    imitation learning (ACT, diffusion, VLA)

  4. 4

    Validate in Sim

    benchmarks before any hardware runs

  5. 5

    Validate on Cell

    staged tests against agreed KPIs

  6. 6

    Deploy & Monitor

    fleet data flows back into the pipeline

  7. fleet data

Deployed on

Bimanual robot cell sorting metal parts into boxes

Bimanual fixed cell

Value today

Two arms, fixed in place. Assembly and pick-and-insert already run productively today.

Mobile manipulator with two arms on a wheeled platform

Mobile manipulator

Accelerating

For machine tending and intralogistics. Certification is under way and maturity is rising fast.

Humanoid robot working at a table alongside a person

Humanoid

Piloting

The most mobile form factor, built for environments made for people.

One pipeline, every form factor: data collected today keeps its value. Teleoperation-based robot learning is proven in production for selected use cases, with hands-on experience in our team and network.

Hackathons and partners

We do not wait for finished answers.

Robot learning is a young field. Much of what will be standard in two years is only just emerging. So we experiment ourselves instead of just reading papers.

Our AI+Robotics hackathons bring the community around one table. Teams of engineers, researchers and students work on real questions: how do you collect data faster and cheaper? How do you get more out of it? Which solution architectures hold up, and what can be built directly on concrete use cases?

Alongside that, we work closely with researchers and startup partners, in Aachen and beyond. Insights from our network inform what we build for our clients.

Collect data fasterUse data betterDesign solution architecturesImplement use cases
Hackathon workspace with robot arms and teams working at their laptops
The AI+Robotics LeRobot Hackathon we hosted with Robotics Collective and IfU
Engineer in a VR headset guiding a robot arm with a hand controller
VR for data collection through teleoperation
Team reviewing object detection results on a screen next to a robot cell
3D computer vision for assembly and disassembly
Engineer running a robot motion plan in simulation on a monitor
Simulation before real world deployment
Two people wiring and testing robot arms on a workbench
Hands-on integration and testing

Insights

From the lab, from hackathons, from projects.

A few clips from our work around robot learning.

Scroll horizontally to browse →

Teleoperation on a Fanuc

An industrial robot on the teleop rig, with our partner TOS.

Bimanual cell, autonomous

Demonstrated first, then the task runs on its own.

Robot learning on a cobot

Learned moves, from grasping to insertion.

Autonomous disassembly

Two arms take a product apart, with our partner Roberto.

Teaching a painting robot

The path comes from a person, not from code.

Teleoperation: clearing a kitchen table

Two arms, one operator. This is how the demonstrations robots learn from are made.