Humanoid robots are attracting growing attention in robotics research. Deploying them in real-world environments, however, remains a significant challenge, despite their considerable potential. Indeed, their form factor allows them to operate in environments designed for humans, such as workshops, dry laboratories, hospitals and homes, and their human-like appearance facilitates acceptance in human-robot interaction. In principle, this enables a wide range of applications in laboratory automation, healthcare and assistive tasks at home, from household chores to elderly care.

This embodiment, however, comes at a high cost. First, safety is a primary concern, as the robot must operate in close proximity to objects that may be fragile. Second, combining locomotion and manipulation is non-trivial: handling heavy objects may compromise the stability of the robot and cause it to fall. Third, perception is more demanding, because the human-like body limits the number and placement of sensors, in contrast to mobile platforms that can be designed around their sensing requirements. Finally, to be of practical use, these robots must solve complex, long-horizon tasks, often involving interaction with people.

These are challenging research problems in their own right, and they also provide a perfect benchmark for reinforcement learning and foundation models. Our objective is the long-term deployment of humanoid robots in real-world environments. To this end, these robots must adapt rapidly and online to their environment, and continue to improve and accumulate knowledge throughout deployment. Within this research theme, we explore a range of techniques, from reinforcement learning to foundation models such as vision-language-action models, as well as their combination.