Post-doctoral Research Visit F/m Post-doc: Learning Contact-rich Locomotion for Humanoid Robots

Le descriptif de l’offre ci-dessous est en Anglais_

**Type de contrat**: CDD

**Niveau de diplôme exigé**: Thèse ou équivalent

**Fonction**: Post-Doctorant

**Contexte et atouts du poste**:
Every year Inria International Relations Department has a few postdoctoral positions **in order to support Inria international collaborations**.

The postdoctoral contract will have a duration of **12 to 24 months**. The default **start date is November 1st, 2025 and not later than January, 1st 2026**. The postdoctoral fellow will be recruited by one of the Inria Centres in France but it is recommended that the time is shared between France and the partner’s country (please note that the postdoctoral fellow has to start his/her contract being in France and that the visits have to respect Inria rules for missions).

The Inria team Hucebot, based in Nancy, France, and University College London (UCL) are collaborating to develop the next generation of control algorithms for humanoid robots thanks to modern artificial intelligence. This partnership is supported by the LEG-AI Inria Associated team, which provides funding for travel and scientific exchanges between the two countries.

The overall goal of this collaboration is to combine forces to mix machine learning algorithms, especially imitation learning and reinforcement learning, with whole-body control, that is, synchronizing all the joints of a humanoid robot to achieve a task while keeping its balance. Both teams have extensive experience with 4-legged (especially UCL) and 2-legged (especially Hucebot) robots, and have invested in several robotic platforms (for quadrupeds: AnyMal, Unitree Go2; for humanoids: Unitree G1, PAL Robotics Talos, IIT iCub). This post-doctoral position will focus on the G1 humanoid robot, which both teams recently acquired, although the other robots (e.g., Talos) can be used for some specific experiments.

This position will require regular stays in London (at least 2 weeks / year) as well as regular remote meetings with the UCL team.

**Mission confiée**:
**Principales activités**:
Humans frequently employ additional contact points to enhance their stability, such as using a handrail or a wall while walking, or to extend their reach, as in grasping a distant object. Similarly, humanoid robots would benefit from a similar strategy. However, current robots prioritize minimizing the number of contacts and utilizing them exclusively for feet and necessary interactions with the environment, like pushing a button [Atkeson 2018]. The few controllers for multi-contact whole-body control assume that the motion is quasi-static, which means that the robot moves much slower than a human [Henze 2016, Rouxel 2025].

One of the main reason for the limited use of contacts is that most of the work in robot control is model-based, that is, using a model of the robot and optimization algorithm to search for the optimal behavior [Henze 2016, Rouxel 2025]. While one can assume that a model of the robot is known, the model of the environment is usually not fully accessible; in particular, some critical quantities like the surface properties have to be guessed. Model-based algorithms also tend to be brittle because they assume a perfect model of the robot, which is never fully accurate.

In the last 5 years, reinforcement learning in simulation revolutionized quadruped locomotion and replaced model-based approaches as the dominating approach [Lee 2020, Hoeller 2024]. By learning robust policies offline, quadrupeds are now capable of highly dynamic motion and take vision/depth sensors as inputs. This learning approach has now replaced model-based approaches as the main paradigm for legged locomotion.

**The objective of this post-doc is to achieve the same transition for humanoid robots: leveraging artificial intelligence to learn policies for biped robots that are highly-dynamic, contact-rich and perception-driven.**

Besides the inherent instability of biped locomotion compared to quadrupeds, one of the specific challenges is that humanoid robot are expected to perform in human-made environments, whereas quadrupeds are more designed for outdoor operation. This means that contacts for humanoids have to be chosen carefully to avoid any potential damage to the environment; for instance, a humanoid robot should not lean on a window or put its hands between fragile objects.

In this post-doc, we will address this challenge with human demonstrations that show examples of additional contacts and combine it with reinforcement learning for learning robust locomotion policies. Both teams, UCL and Hucebot have extensive experience in learning algorithms for robots. The post-doc will rely on the recent work of the team about legged robot control [Turisi 2024, Penco 2019], multi-contact whole-body control [Rouxel 2024a, Totsila 2024] and learning from demonstration with flow matching algorithms [Rouxel 2024b], as well as the expertise of UCL for robot vision [Liu 2024]. The