Objectives of the chair
Robotics has seen tremendous progress in recent years thanks to advances in simulators, optimizers, and reinforcement learning. Despite impressive demonstrations, however, these methods have yet to be fully deployed in real-world settings, and scaling beyond the lab remains a challenge.
What if we could harness the full potential of constrained optimization and reinforcement learning to develop safe and efficient policies for dynamic locomotion and manipulation tasks? In this chair, we propose to achieve this goal by revisiting the design of reinforcement learning algorithms through the lens of optimization, leveraging expertise gained through years of progress in model-predictive control and recent results in parallel simulators and optimizers.
Our goal is to achieve faster and more accurate convergence by reducing the computational burden and providing strict guarantees on the resulting optimal policy. We will then incorporate global strategies to escape local minima and enable planning across various contact modes (e.g., due to intermittent contact interactions). We will further explore the integration of multimodal sensory information, including force and touch detection, leveraging recent results with foundation models. These enhancements will increase the robustness and safety of our policies during physical interactions, enabling reliable robotic manipulation and locomotion in complex environments.
Our team’s unique experimental capabilities and extensive expertise in reinforcement learning and optimal control will enable the establishment of a comprehensive framework for the optimization and control of robotics policies, ultimately enabling the development of robots capable of complex, adaptive, and reliable movements deployed in real-world applications.
The chair will benefit from strong interactions with the synergy chair C3PO, as well as other chairs in numerical optimization and machine learning. Results will be translated into practical problems through an effective collaboration with the robotics manufacturer PAL and the end-user AIRBUS. The project will also consider direct socioeconomic impacts, with a particular focus on direct outreach to students through the creation of an international exchange program and to young audiences through specific robotics activities.
Research objectives
Unified theoretical framework for reinforcement learning and model predictive control Apprentissage par renforcement différentiable avec des modes mixtes discrets-continus. Multi-modal sensing for partially observed systems (including haptic feedback) Safe and efficient policy optimization for dynamic robotic behaviors (locomotion and manipulation)
Industrial applications
- The chair will deploy research results on their specific application needs, leveraging already existing collaborations with N. Mansard
- Toward Force Feedback in Model-Predictive Control, PhD thesis of S. Kleff at NYU, co-Supervised by L. Righetti and N. Mansard,
- C3PO, chaire ANITI – projet de thèse de doctorat en cotutelle encadrée par L. Righetti, N. Mansard et O. Stasse
- International Student Exchange Program, M.Sc. students from NYU and the Toulouse region will be able to conduct research for a duration of 6 months in the laboratory of the other country
- Educational outreach, students from diverse socio-economic backgrounds, including an international robot programming competition with middle schools from New York City and the Toulouse region.
Main scientific goals
- A comprehensive algorithmic framework for scalable, complex and reliable robotic behaviors
- Demonstration of novel manipulation and locomotion behaviors on real robots (manipulators, quadrupeds and humanoids)