Conventional aquatic robots are typically constrained by fixed morphology and single-mode locomotion, limiting adaptability to unstructured environments. Inspired by the diverse fin-driven locomotion strategies of natural fish, we present a self-reconfigurable robotic fish swarm system capable of operating across a wide range of aquatic conditions and tasks. Each robotic fish unit features autonomous physical assembly and disassembly capabilities by electropermanent magnets, enabling both connection and intermodule communication, allowing the swarm to dynamically reconfigure its morphology. Systematic evaluation reveals that swarm configurations substantially outperform individual units in various locomotion performance, including stability, maneuverability, swimming speed, energy efficiency, and multimodal locomotion ability. The swarm also demonstrates collaborative capabilities in navigating through complex environments, manipulating obstacles, and transporting objects, both in laboratory and outdoor aquatic settings. This work provides a framework for aquatic robots to adapt to unstructured environments and complex tasks, contributing to the advancement of multifunctional, reconfigurable robotic systems.
@article{si2026selfreconfigurable,title={Self-reconfigurable robotic fish swarms: Collective achievement of diverse locomotion and challenging aquatic tasks},journal={Science Advances},volume={12},number={2},pages={eadz2458},year={2026},publisher={American Association for the Advancement of Science},doi={10.1126/sciadv.adz2458},}
SPL
The Optimal Condition Number for Leaky ReLU Function
The condition number of a network layer, the ratio of its optimal upper and lower Lipschitz constants, measures how severely the layer distorts distances between inputs. Extending the ReLU result of Xia and Zhou, we establish the analogous pair of results for Leaky ReLU with leakage coefficient between 0 and 1. A pointwise algebraic identity decomposes the squared distance between Leaky ReLU images into two standard ReLU terms and one linear term, so existing concentration estimates transfer without modification. The optimal condition number equals sqrt(2(1+alpha^2))/(1+alpha): no choice of weights improves on it, Gaussian weights achieve it asymptotically, and it decreases strictly from sqrt(2) at alpha = 0 to 1 at alpha = 1, quantifying the geometric benefit of leakage.
@article{chang2026leakyrelu,title={The Optimal Condition Number for Leaky ReLU Function},journal={IEEE Signal Processing Letters},year={2026},note={Accepted}}
L-CSS
Measure-Based Razumikhin Conditions for Delay Systems
In integral Lyapunov-Razumikhin conditions, the current value of a Lyapunov function is compared with a weighted integral of its recent history instead of the maximum over the delay window. This paper formulates the comparison with an arbitrary finite positive Borel measure on the delay window and establishes global stability, uniform global asymptotic stability, and input-to-state stability by an elementary record-point argument, recovering the known integral-kernel conditions as the absolutely continuous case. All finite Lp-type formulations are shown to be equivalent to the integral case under an exact reparameterization of the gains, so the exponent carries no structural information; for a scalar linear system with one discrete delay a certificate exists if and only if the measure places positive mass at the delay. The p = infinity endpoint, governed by the support of the measure, connects the framework to the classical maximum-type condition.
@unpublished{chang2026razumikhin,title={Measure-Based Razumikhin Conditions for Delay Systems},year={2026},note={IEEE Control Systems Letters, revise and resubmit}}
NeurIPS-W
Projected Hessian Direction Estimation by Comparisons under Hölder Smoothness
In 18th International Workshop on Optimization for Machine Learning (OPT), NeurIPS, 2026
@inproceedings{chang2026opt,title={Projected Hessian Direction Estimation by Comparisons under H{\"o}lder Smoothness},booktitle={18th International Workshop on Optimization for Machine Learning (OPT), NeurIPS},year={2026},note={Poster}}
NeurIPS-W
Sharpness of Transport Bounds on Rényi Divergence on Curved Spaces
In Workshop on Geometric Distributional Deep Learning (GDDL), NeurIPS, 2026
@inproceedings{chang2026gddl,title={Sharpness of Transport Bounds on Rényi Divergence on Curved Spaces},booktitle={Workshop on Geometric Distributional Deep Learning (GDDL), NeurIPS},year={2026},note={Poster}}
NeurIPS-W
Sharpness of Transport Bounds on Rényi Divergence along Diffusions
In Workshop on AI for Stochastic Dynamics (STODY), NeurIPS, 2026
@inproceedings{chang2026stody,title={Sharpness of Transport Bounds on Rényi Divergence along Diffusions},booktitle={Workshop on AI for Stochastic Dynamics (STODY), NeurIPS},year={2026},note={Poster}}
NeurIPS-W
The Newsvendor with Distributional Advice
In Second Workshop on Machine Learning and Operations Research (ML×OR), NeurIPS, 2026
@inproceedings{chang2026mlxor,title={The Newsvendor with Distributional Advice},booktitle={Second Workshop on Machine Learning and Operations Research (ML×OR), NeurIPS},year={2026},note={Poster}}
APS DFD
A mean-field game on the diffusive point-vortex system and short-time solutions of the two-dimensional Navier–Stokes equations
In 79th Annual Meeting of the APS Division of Fluid Dynamics, 2026
@inproceedings{chang2026vortex,title={A mean-field game on the diffusive point-vortex system and short-time solutions of the two-dimensional Navier--Stokes equations},booktitle={79th Annual Meeting of the APS Division of Fluid Dynamics},year={2026},note={Accepted abstract}}
IROS-W
Spontaneous Directional Order in One-Bit Extremum-Seeking Swarms
In IROS 2026 Workshop on Miniature Multiterrain and Multimodal Locomotion: from Biology to Robotics, 2026
@inproceedings{chang2026directional,title={Spontaneous Directional Order in One-Bit Extremum-Seeking Swarms},booktitle={IROS 2026 Workshop on Miniature Multiterrain and Multimodal Locomotion: from Biology to Robotics},year={2026},note={Poster}}
IROS-W
Which Module Shapes Can Pivot?
In IROS 2026 Workshop: Challenges and Applications Prospects for Reconfigurable Modular Robots, 2026
@inproceedings{chang2026pivot,title={Which Module Shapes Can Pivot?},booktitle={IROS 2026 Workshop: Challenges and Applications Prospects for Reconfigurable Modular Robots},year={2026}}
IROS-LBR
Geometric Hoverability Analysis of Generalized Tilted Multirotors
In 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
@inproceedings{chang2026tilted,title={Geometric Hoverability Analysis of Generalized Tilted Multirotors},booktitle={2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},year={2026},note={Late Breaking Results poster}}
ICRA-W
A Geometric Method of Generalized Hoverability and Robustness Analysis for Marine Inspection UAVs
In ICRA 2026 Workshop on Aerial Inspection for Marine Infrastructures, 2026
@inproceedings{chang2026hoverability,title={A Geometric Method of Generalized Hoverability and Robustness Analysis for Marine Inspection UAVs},booktitle={ICRA 2026 Workshop on Aerial Inspection for Marine Infrastructures},year={2026},note={Poster}}
RA-L
Geometric Classification of Singularity Varieties for (2n+1)R Circular Manipulators
2026
Under review at IEEE Robotics and Automation Letters
The singular configurations of a redundant manipulator form the organizing skeleton of its kinematics. For the (2n+1)R circular manipulators every three consecutive revolute axes intersect at a joint center, and these configurations form an algebraic variety in the tangent-half-angle parameters. This paper gives a complete geometric and algebraic classification of the singularity variety for every n >= 3 with positive link offsets. A reciprocal-line reduction forces every screw reciprocal to all joint axes to be the Pluecker vector of a line through the bent centers; a projective incidence trichotomy propagates this line along the chain and yields the classification theorem. The radical singularity ideal coincides with an explicit combinatorial ideal, the irreducible components are coordinate-affine linear subspaces organized into five parametrized families, and the exact component count N_n = (n-1)(2n-3) follows, establishing a formula previously conjectured from finite computational evidence.
@unpublished{chang2026singularity,title={Geometric Classification of Singularity Varieties for (2n+1)R Circular Manipulators},year={2026},note={Under review at IEEE Robotics and Automation Letters}}
SPL
A Phase Transition for Shifted-Composition Transport Bounds on Rényi Divergence
@unpublished{chang2026renyi,title={A Phase Transition for Shifted-Composition Transport Bounds on Rényi Divergence},year={2026},note={Under review at IEEE Signal Processing Letters}}
@unpublished{chang2026bookmaking,title={Optimal Online Bookmaking on Bipartite Graphs},year={2026},note={Under review at IEEE Control Systems Letters}}
L-CSS
One-Sided Budget Uncertainty in General Lotto Games
@unpublished{chang2026lotto,title={One-Sided Budget Uncertainty in General Lotto Games},year={2026},note={Under review at IEEE Control Systems Letters}}