Open-source software helped accelerate the development of artificial intelligence. Pepijn Kooijmans of Hugging Face believes robotics could now be approaching a similar turning point. Speaking at Future Pulse, he argued that cheaper hardware, shared datasets and open robot-learning models are lowering the barriers to experimenting with physical AI.
Robotics has traditionally been a relatively closed and expensive field. Industrial robots are often programmed for narrowly defined tasks, while the data, software and hardware needed to develop more flexible systems are spread across manufacturers, research institutions and factories.
That model is beginning to change. According to Pepijn Kooijmans, robot lead at Hugging Face’s LeRobot project, robotics is currently going through what he describes as its “open-source moment”. The combination of imitation learning, shared datasets and increasingly capable open models is making it possible for a much broader community to experiment with robot learning.
Kooijmans traced that development back to his own experiments several years ago. At the time, research projects such as Google’s RT-1 and RT-2 were demonstrating how robots could learn from large amounts of demonstrations rather than relying exclusively on conventional programming.
Using a small, partly 3D-printed robot arm and an early imitation-learning model, he recorded around thirty demonstrations of a simple pick-and-place task. Instead of programming trajectories and positions individually, he demonstrated the task and trained the model on those examples. The result was a system that could reproduce the behaviour without being explicitly programmed for every movement.
For Kooijmans, that represented an important change in how robots could be developed. Rather than programming every task separately, developers can collect observations and actions and train models to reproduce human behaviour. Camera images, robot joint positions and instructions can be combined by vision-language-action models, which subsequently predict the actions a robot should perform.
From code to data
That shift also changes what becomes valuable in robotics.
“The tasks are moving from code into data,” was one of Kooijmans’ central messages at Future Pulse. Instead of describing every step of a task in software, developers increasingly need examples showing the robot what successful execution looks like.
LeRobot is intended to provide an open environment for that approach. The project brings together software for connecting robots, teleoperating them, recording demonstrations, storing datasets, training models and deploying those models back onto robots.
One of the goals is to reduce the cost of getting started. LeRobot has therefore supported inexpensive robot arms including the SO-100 and SO-101. Kooijmans said systems of this type can be built for roughly €100 using 3D-printed components, allowing developers to experiment with many of the same learning techniques being applied to substantially more expensive robotic platforms. The software can also be connected to other systems, including humanoid robots.
Teleoperation forms another part of the process. A person can control a robot using VR equipment or a smaller “leader” arm while the target robot follows the movement. The resulting camera images, states and actions are stored as demonstrations that can subsequently be used for training.
Standardising how that information is stored is important because robot data remains comparatively scarce and fragmented. Unlike language models, which can be trained on enormous quantities of text available online, robotics has no equivalent of the public internet containing large volumes of labelled physical interactions.
Factories and robot suppliers already generate significant amounts of operational data, but much of it is inaccessible to outside researchers. LeRobot therefore uses a common dataset format intended to make robot data easier to share and reuse across projects.
An open recipe for robot learning
Kooijmans illustrated the approach with a project in which a robot learned to flatten and fold T-shirts. Folding deformable objects is difficult to solve through conventional robot programming because the starting configuration can vary substantially.
The project used almost 6,000 recorded episodes involving more than 25 different T-shirts, several operators and different configurations. Rather than presenting the resulting demonstration as a solved problem, Kooijmans emphasised the value of publishing the entire development process.
An open “recipe” allows another team to start with an existing method, apply it to another task and improve it rather than rebuilding the complete system from scratch.
That principle is also visible in the growing community around LeRobot. Kooijmans said almost 80,000 robotics datasets had been shared through Hugging Face, while the project had accumulated around 27,000 GitHub stars and a community of more than 20,000 people across platforms such as Discord and GitHub at the time of his presentation.
He sees the mix of researchers, companies and individual developers as particularly important. Industrial companies bring concrete automation problems, while researchers can develop new methods and feed those results back into practical applications.
From VLA to world-action models
The underlying technology is also evolving.
Vision-language-action models are increasingly being trained on data from multiple robots and larger collections of demonstrations. Some models can already transfer elements of learned behaviour between different robot embodiments.
Kooijmans also pointed to the emergence of world-action models. Instead of predicting only the robot’s next action, these systems attempt to model what will happen in the environment as well. One approach is to predict future video, giving the model an additional learning signal about how the physical world changes following an action.
Another development could eventually make demonstration even more intuitive. Rather than controlling the robot directly through teleoperation, a person could potentially demonstrate a task by performing it themselves on video. A robot model could then use that example to infer the intended task.
Kooijmans asked the audience to imagine an experienced factory worker showing a robot how a task is performed, after which the behaviour could be deployed across multiple machines. Such an approach could be particularly relevant to high-mix manufacturing, where traditional automation can become expensive when products and processes frequently change.
Open source also raises questions
The development is still at an early stage.
Asked whether he had personally seen general-purpose robot policies running in normal production rather than pilots, Kooijmans said he had not. Robotics remains difficult because every part of the physical system has to work together reliably, and errors in perception, models, hardware or control can accumulate.
Hardware differences are another unresolved problem. A policy trained on one robot does not automatically transfer perfectly when motors, joint ranges or mechanical components change. Researchers are working on more general representations that could make data from different robot types easier to combine, but Kooijmans acknowledged that the field has not yet reached the standardisation seen in language modelling.
Safety was also raised during the discussion. Kooijmans pointed to work on alignment and steerability: ensuring that a robot’s behaviour remains consistent with the human instruction and intended outcome. The question becomes increasingly important as robots move from fixed programs toward systems capable of adapting their behaviour from data.
For Kooijmans, these uncertainties are another argument for keeping a significant part of the robotics stack open. Open models can be inspected, modified and, where necessary, rolled back. Companies can also adapt them to proprietary processes rather than depending entirely on intelligence controlled by an external supplier.
The comparison with large language models is deliberate. Open language models have narrowed the gap with proprietary systems while allowing organisations to run, inspect and customise models themselves. Kooijmans believes robotics could follow a similar path.
If that happens, the important question will not only be who manufactures the robots, but also who owns and controls the intelligence running them.
His message to the Future Pulse audience was therefore practical rather than futuristic: begin experimenting now. Start with a small robot and a narrowly defined task, collect a few dozen demonstrations and see what the current generation of open tools can already do.
For countries such as the Netherlands, with established expertise in machine building, system integration and specialised manufacturing, Kooijmans sees an opportunity to participate in the development of robot intelligence rather than importing it later as a finished technology.
Open source will not remove the remaining technical problems in robot learning. But by making hardware designs, datasets, models and development methods available to a broader community, it could substantially increase the number of people able to work on them.
