Artificial Intelligence How robots learn from few data

Related Vendors

FAULHABER Drive Systems

Wuxi InfiMotion Technology Co., Ltd.

If robots are to be beneficial, they need to get their intended tasks described as precisely as possible. This usually requires large amounts of data. Can this also be done more efficiently?

Sven Behnke heads the Computer Science VI - Intelligent Systems and Robotics department and the Autonomous Intelligent Systems working group at the University of Bonn. He is also a member of the Learning Capable Robotics working group of the Learning Systems platform.

Robots are of great use in industrial mass production. Without industrial robots taking over repetitive tasks, no car would be produced in Germany anymore. Mobile robots, for example, transport shelves in warehouses or meals in hospitals. Simple robots already help in the household, for instance with floor cleaning or lawn care.

Currently, a close task definition is required for robots to be useful, for example "move an object from A to B". The deployment environment also needs to be structured, for example by providing the object at a known location in a known position. Meanwhile, research is working on new areas of application for robots: in the future, they are expected to work directly with people in production, assist people in need of assistance in everyday life or support emergency services in coping with disasters. However, these open, complex application domains require more cognitive skills than current autonomous robots have. Today, remotely controlled robots can solve numerous tasks in complex environments with the human intelligence of their operator. In this case, teleoperation places a human into an avatar robot. The human operator can easily cope with new situations and can flexibly transfer their existing knowledge to the current conditions. They recognize problems with the execution and quickly develop alternatives for action.

How can we equip robots with cognitive abilities?

In recent years, impressive success has been achieved in related fields through deep learning, for example in visual perception, speech recognition and synthesis, as well as in dialogue systems like Chat GPT. These are based on training large models with gigantic amounts of data. Such basic models capture extensive world knowledge and can quickly adapt to specific tasks, for example through transfer learning or in-context learning. So how can we continue this success story for robotics? The first steps in this direction are multimodal models, which are trained not only with one modality, i.e. exclusively texts, images or speech, but with data from several modalities, such as CLIP from Open AI. Even if the collection of real robotic interaction data is complex, there are initiatives to merge data from different robots and tasks, such as Open X-Embodiment. Models trained with this can solve a variety of manipulation tasks better than models that were only trained with specific data.

Another possibility is to generate interactions in a simulation. The challenges here are to make the simulation realistic and to transfer what has been learned in the simulation to reality, also known as the Sim2Real gap.

Learning from large amounts of data in robotics as well?

The human model shows us that data-efficient learning is possible. Specific learning models are required for this, which have been optimized evolutionarily and require little data through the use of prior knowledge - keyword inductive bias. While we can't learn any task anymore, we do learn the tasks that life sets us faster and better.

In order to achieve comparable data efficiency in robots, similar learning models are needed that have a suitable inductive bias. In my view, it is helpful to orientate towards the structure of the human cognitive system. In particular, robots need not only a fast, parallel sensorimotor system 1 for routine tasks, but also a system 2 for higher cognitive functions, such as planning or assessing their own limits.

With an appropriate cognitive architecture, teleoperation offers great opportunities to gradually transfer human competencies to autonomous functions and thus increasingly less dependent on humans as operators.