Institution profile

Ecole Centrale de Nantes

Academic institutioneurope · fr
Official website
Research library24linked papers
Opportunities0open roles
Selected work

Representative Papers

The Twist Decomposition of Serial Robots Under Lower-Mobility Tasks

Jul 21, 2026

This work addresses the challenges of kinematic redundancy and task-space decoupling in serial manipulators performing low-degree-of-freedom tasks. To overcome these issues, the authors propose a geometrically defined screw projector that directly decomposes the end-effector twist into task-relevant and redundant components, thereby establishing a compact inverse kinematics framework. Unlike conventional approaches relying on Jacobian null-space projection, this method leverages geometric screw decomposition to intuitively separate motions inside and outside the task space, offering a unified treatment of both kinematic and task redundancy. Experimental results demonstrate that the proposed approach enables efficient, intuitive, and natural motion control while effectively managing redundancy.

0 citationsRead paper

Can Transformers Learn to Verify During Backtracking Search?

May 21, 2026

This work addresses the susceptibility of Transformer-based models to historical trajectory interference in backtracking search tasks, which undermines their ability to make consistent decisions based solely on the current state. To mitigate this issue, the authors propose Selective State Attention (SSA), a novel mechanism that employs structured attention masks and trace-level localization to constrain decoder-style Transformers to reason exclusively from the current search state—without requiring modifications to training data or objectives. Evaluated across diverse domains including 3-SAT, graph coloring, Blocks World, and backtracking parsing, SSA consistently produces identical outputs for identical states despite varying histories, significantly outperforming baseline models and demonstrating its efficacy in enhancing state-dependent decision-making capabilities.

0 citationsRead paper

Natural gradient descent with momentum

Apr 16, 2026

This work addresses the challenge of optimizing complex loss functions—such as Kullback–Leibler divergence or PDE residuals—over nonlinear manifolds like neural or tensor networks, where conventional and natural gradient descent often converge to suboptimal local minima with inefficient update directions. The paper introduces, for the first time, a momentum-augmented variant of natural gradient descent that integrates Heavy-Ball and Nesterov-type inertial mechanisms into a manifold-aware optimization framework. By leveraging tangent-space projections and Gram matrix preconditioning, the method achieves momentum-driven, locally optimal updates directly in function space. Empirical results demonstrate substantial improvements in convergence behavior for tasks including density estimation and physics-informed learning, effectively mitigating poor local minima and enhancing the quality of optimization trajectories.

0 citationsRead paper
Recent publications

Latest Papers

The Twist Decomposition of Serial Robots Under Lower-Mobility Tasks

Jul 21, 2026

This work addresses the challenges of kinematic redundancy and task-space decoupling in serial manipulators performing low-degree-of-freedom tasks. To overcome these issues, the authors propose a geometrically defined screw projector that directly decomposes the end-effector twist into task-relevant and redundant components, thereby establishing a compact inverse kinematics framework. Unlike conventional approaches relying on Jacobian null-space projection, this method leverages geometric screw decomposition to intuitively separate motions inside and outside the task space, offering a unified treatment of both kinematic and task redundancy. Experimental results demonstrate that the proposed approach enables efficient, intuitive, and natural motion control while effectively managing redundancy.

0 citationsRead paper

Can Transformers Learn to Verify During Backtracking Search?

May 21, 2026

This work addresses the susceptibility of Transformer-based models to historical trajectory interference in backtracking search tasks, which undermines their ability to make consistent decisions based solely on the current state. To mitigate this issue, the authors propose Selective State Attention (SSA), a novel mechanism that employs structured attention masks and trace-level localization to constrain decoder-style Transformers to reason exclusively from the current search state—without requiring modifications to training data or objectives. Evaluated across diverse domains including 3-SAT, graph coloring, Blocks World, and backtracking parsing, SSA consistently produces identical outputs for identical states despite varying histories, significantly outperforming baseline models and demonstrating its efficacy in enhancing state-dependent decision-making capabilities.

0 citationsRead paper

Natural gradient descent with momentum

Apr 16, 2026

This work addresses the challenge of optimizing complex loss functions—such as Kullback–Leibler divergence or PDE residuals—over nonlinear manifolds like neural or tensor networks, where conventional and natural gradient descent often converge to suboptimal local minima with inefficient update directions. The paper introduces, for the first time, a momentum-augmented variant of natural gradient descent that integrates Heavy-Ball and Nesterov-type inertial mechanisms into a manifold-aware optimization framework. By leveraging tangent-space projections and Gram matrix preconditioning, the method achieves momentum-driven, locally optimal updates directly in function space. Empirical results demonstrate substantial improvements in convergence behavior for tasks including density estimation and physics-informed learning, effectively mitigating poor local minima and enhancing the quality of optimization trajectories.

0 citationsRead paper