Authors: Ale, Côme, Constant et al

By Sébastien Kleff (1,2,3), Armand Jordana (1,2,4), Nicolas Mansard (3,5), Ludovic Righetti (1,2,5)

Model-predictive control is an appealing framework to control robots due to its ability to exploit both sensory information and model predictions. But its performance remains fundamentally limited in tasks involving contact with the environment, in part because optimal control policies do not reason over force measurements. In this article, we propose a first complete answer to this issue by introducing a novel approach that systematically includes measured efforts into the optimal control loop. We propose to augment the state-space with a visco-elastic model of the contact force in the task space. We derive a complete predictive controller with an efficient formulation whose implementation is released in open-source. We conduct extensive comparisons with two other methods: the classical model-predictive control formulation, which inherently restricts the feedback to position and velocity information, and our previous approach that enabled torque feedback in the joint space. We demonstrate through simulation studies and hardware experiments, the benefit of exploiting Cartesian force measurements in the model-predictive control framework to achieve challenging contact tasks.

Project page: https://gepettoweb.laas.fr/articles/kleff_tro_2025.html

^1^ NYU - New York University [New York]

By Armand Jordana (1,2), Sébastien Kleff (2,4,6), Arthur Haffemayer (4,5), Joaquim Ortiz-Haro (1), Justin Carpentier (3), Nicolas Mansard (4,5), Ludovic Righetti (2)

Model Predictive Control has emerged as a popular tool for robots to generate complex motions. However, the real-time requirement has limited the use of hard constraints and large preview horizons, which are necessary to ensure safety and stability. In practice, practitioners have to carefully design cost functions that can imitate an infinite horizon formulation, which is tedious and often results in local minima. In this work, we study how to approximate the infinite horizon value function of constrained optimal control problems with neural networks using value iteration and trajectory optimization. Furthermore, we demonstrate how using this value function approximation as a terminal cost provides global stability to the model predictive controller. The approach is validated on two toy problems and a real-world scenario with online obstacle avoidance on an industrial manipulator where the value function is conditioned to the goal and obstacle.

Project page: https://gepettoweb.laas.fr/articles/jordana__value_2025.html

^1^ NYU - New York University [New York]

By Armand Jordana (1), Avadesh Meduri (2), Etienne Arlaud (3), Justin Carpentier (3), Ludovic Righetti (1,4)

In robotics, designing robust algorithms in the face of estimation uncertainty is a challenging task. Indeed, controllers often do not consider the estimation uncertainty and only rely on the most likely estimated state. Consequently, sudden changes in the environment or the robot's dynamics can lead to catastrophic behaviors. In this work, we present a risk-sensitive Extended Kalman Filter that allows doing outputfeedback Model Predictive Control (MPC) safely. This filter adapts its estimation to the control objective. By taking a pessimistic estimate concerning the value function resulting from the MPC controller, the filter provides increased robustness to the controller in phases of uncertainty as compared to a standard Extended Kalman Filter (EKF). Moreover, the filter has the same complexity as an EKF, so that it can be used for real-time model-predictive control. The paper evaluates the risk-sensitive behavior of the proposed filter when used in a nonlinear model-predictive control loop on a planar drone and industrial manipulator in simulation, as well as on an external force estimation task on a real quadruped robot. These experiments demonstrate the abilities of the approach to improve performance in the face of uncertainties significantly.

Project page: https://gepettoweb.laas.fr/articles/jordana__ekf_icra2025.html

^1 NYU Tandon School of Engineering

Published in IEEE International Conference on Robotics and Automation (ICRA)

Nonlinear model-predictive control has recently shown its practicability in robotics. However it remains limited in contact interaction tasks due to its inability to leverage sensed efforts. In this work, we propose a novel model-predictive control approach that incorporates direct feedback from force sensors while circumventing explicit modeling of the contact force evolution. Our approach is based on the online estimation of the discrepancy between the force predicted by the dynamics model and force measurements, combined with high-frequency nonlinear model-predictive control. We report an experimental validation on a torque-controlled manipulator in challenging tasks for which accurate force tracking is necessary. We show that a simple reformulation of the optimal control problem combined with standard estimation tools enables to achieve state-of-the-art performance in force control while preserving the benefits of model-predictive control, thereby outperforming traditional force control techniques. This work paves the way toward a more systematic integration of force sensors in model predictive control.

Project page: https://gepettoweb.laas.fr/articles/jordana__icra2024.html

^1 NYU - New York University [New York]

Submitted to IEEE Transactions on Robotics (T-RO)

The promise of model-predictive control in robotics has led to extensive development of efficient numerical optimal control solvers in line with differential dynamic programming because it exploits the sparsity induced by time. In this work, we argue that this effervescence has hidden the fact that sparsity can be equally exploited by standard nonlinear optimization. In particular, we show how a tailored implementation of sequential quadratic programming achieves state-of-the-art model-predictive control. Then, we clarify the connections between popular algorithms from the robotics community and well-established optimization techniques. Further, the sequential quadratic program formulation naturally encompasses the constrained case, a notoriously difficult problem in the robotics community. Specifically, we show that it only requires a sparsity-exploiting implementation of a state-of-the-art quadratic programming solver. We illustrate the validity of this approach in a comparative study and experiments on a torque-controlled manipulator. To the best of our knowledge, this is the first demonstration of nonlinear model-predictive control with arbitrary constraints on real hardware.

Project page: https://gepettoweb.laas.fr/articles/jordana_kleff_meduri__sqp_tro_2025.html

Work of 8 young apprentice researchers (10-16 yo) on a sunny July afternoon

Recent technological progress is supported by the generalization of numerical tools for simulating, controlling, and observing physical systems. Yet, by focusing on more and more complex phenomena, our conventional tools are falling short of meeting the growing expectations of engineers, whether in terms of accuracy or computation time. Data-driven approaches, in particular neural networks, offer promising alternatives to address these new challenges. These models can capture complex, nonlinear relationships in physical systems, and shift the burden from manual derivation of tedious mathematical formulas towards large-scale data collection. However, these methods often sacrifice stability, robustness, precision, and more generally guarantees classically offered by traditional approaches. In this thesis, we propose combining the fields of physics, deep learning, and control theory to propose new hybrid methods, taking advantage of the expressivity of neural networks, while relying on inductive biases from physics. We describe theoretical tools (discussed in Part 1) related to the simulation of dynamical systems and connect them to neural network design. In a second time (Part 2), we leverage these insights to design control algorithms and simulation techniques addressing the resolution of complex problems related to partial differential equations. Finally, in Part 3, we focus on larger-scale simulations such as fluid dynamics and counterfactual reasoning. Our work has been presented at scientific conferences in the field of artificial intelligence and control theory. By bridging the gap between physics and machine learning, we believe that this paves the way toward a new generation of methods for the simulation and control of physical systems.

About the speaker: Steeven Janny just completed his PhD at INSA Lyon within the LIRIS and LAGEPP research labs, specializing in the interdisciplinary realms of physics, control, and machine learning. He harbors a profound interest in the fusion of physical simulation with deep learning techniques, and are eager to further explore related areas such as ML-based control and robotics. With a solid foundation in electrical engineering and signal processing, complemented by a degree in artificial intelligence, Steeven brings a diverse skill set to his research pursuits. In addition to his academic endeavors, Steeven Janny is committed to teaching and is a certified "agrégé" teacher in engineering and computer science, ensuring that education remains an integral part of their career trajectory.