LEG-AI
Learning and Generative AI methods for Control of Legged Robots
Building on a longstanding track record of previous joint publications and projects, the strategic collaboration LEG-AI between the Inria HUCEBOT team and the robotics & AI researchers of the Department of Computer Science of the University College London (UCL) seeks to push the boundaries of legged locomotion and develop adaptive, autonomous control strategies for both quadruped and humanoid robots in challenging environments.
Established in 2025, this collaboration which is driven by the use of state-of-the-art methods in deep learning, generative AI, vision transformers, and evolutionary algorithms, promotes knowledge sharing, methodologically and experimentally, and fosters collaboration. Such results are reached through a structured program of research visits between the centre Inria de l’Université de Lorraine and UCL which enhance the condition to share the teams’s specialized expertise and strengthen experimental evaluation by sharing complementary robotics platforms.
CONTEXT
Robotics is currently undergoing a significant transformation, where classical model-based and optimization-based control methods are increasingly being complemented or even partially replaced by techniques developed in the machine learning community. These advancements include the use of reinforcement learning to develop robust, reactive policies from large-scale simulations, environment awareness via exteroceptive state estimation, and processing of raw pixel input through deep learning encoders, replacing traditional object detection and pose estimation techniques. These innovations are significantly expanding the capabilities of robotic systems, driving unprecedented levels of autonomy, adaptability, and efficiency in complex and dynamic environments.
OBJECTIVES
The creation of this joint team is driven by their common objectives which are:
- Advancing humanoid robot motion generation
- Enhancing Inria’s humanoid robots’ locomotion capabilities on uneven terrains
- Advancing locomotion capabilities in humanoid robots
- Developing groundbreaking approaches for legged robot motion generation
ADDED VALUE
The collaboration between UCL and Inria leverages the unique strengths and complementary expertise of both institutions, creating a synergistic partnership that significantly enhances the research capabilities of each partner.
Department of Computer Science
University College London (UCL)
UCL brings extensive proficiency in computer vision, and advanced reinforcement learning techniques tailored for quadruped locomotion and manipulation tasks. This expertise enables the development of sophisticated perception systems and adaptive control strategies that are crucial for expanding humanoids’ capabilities to adapt to complex environments while generating human-like movements. Additionally, UCL’s experience in data-driven path planning and deep learning for vision-based applications provides robust methodologies for integrating raw sensory inputs into effective motion generation frameworks.
HUCEBOT research team
Centre Inria de l’Université de Lorraine
Inria, on the other hand, offers deep expertise in humanoid robot control, including the development of multi-contact motion controllers and imitation learning methods. Inria’s strengths in human-robot interaction and teleoperation further complement UCL’s capabilities, facilitating the creation of more intuitive and responsive humanoid motion strategies. The availability of Inria’s humanoid platforms, such as PAL Talos and Unitree G1, alongside UCL’s quadruped robots like Boston Dynamics Spot and Unitree Go1, provides a diverse and versatile testing environment that fosters innovative cross-platform solutions.
Furthermore, this collaboration creates a unique opportunity for cross-pollination between cognitive applications developed for quadruped robots and their application in the context of humanoid robotics. Research primarily focused on cognitive strategies and high-level planning for quadrupeds can be effectively transferred to enhance the intelligence and adaptability of humanoid robots. Conversely, Inria’s expertise in low-level controllers for multi-contact tasks, as well as their knowledge in human-robot collaboration and teleoperation, can significantly benefit quadrupedal robots, where the emphasis has traditionally been on navigation capabilities. By integrating cognitive planning from quadrupeds with robust multi-contact control and intuitive human-robot interaction techniques from Inria, the joint team can develop more versatile and capable robotic systems.
Joint Team Members
HUCEBOT – Centre Inria de l’Université de Lorraine
Serena Ivaldi, DR2
Jean-Baptiste Mouret, DR1
Enrico Mingo Hoffman, ISFP
Dionis Totsila, PhD student
Constantinos Tsakonas, PhD student Inria
Ioannis Tsikelis, PhD student
Robotics and AI researchers of the Department of Computer Science of the University College London
Valerio Modugno, Associate Lecturer
Dimitrios Kanoulas, Professor
