arXiv subjects

Yixuan Feng

Publications and source records attributed to Yixuan Feng.

4 recordsLinked to original sources

arXiv2026

Robotic Multiphase Interaction: Manipulating Coupled Liquid and Solid Dynamics with a World Model

This work presents \textit{Robotic Multiphase Interaction (RMI)}, a setting in which liquid enters a porous material and interacts mechanically with its deforming solid skeleton. Manipulation can therefore change pore volume, expel or redistribute retained liquid, and alter grasp stability at the same time. Spilled liquid can also create safety risks in domestic and manufacturing settings. This differs from most manipulation of solid objects and from tasks that involve both liquid and solid while keeping the phases spatially separate. We study a sponge filled with water as the first RMI example. We use implicit incompressible porous flow with smoothed particle hydrodynamics as the dynamics engine and enable robotic manipulation by adding Coulomb contact memory, hybrid velocity and force regulation, and a stability gate for lifting. The resulting environment connects robot commands to changes in the coupled liquid and solid state. A world model conditioned on actions predicts how this state evolves under candidate commands, while a temporal UNet generates actions using either Diffusion Policy or rectified flow matching. Our world model reduces retained water prediction error by more than $60\%$ compared with the baseline. The best action sequence selected by the world model from policy proposals further reduces the predicted terminal water error by about half. These improvements show that modelling the coupled liquid and solid state helps the robot predict how its actions affect both the porous object and the liquid held inside.

cs.RO↗

arXiv2025

The effects of the Gravitational Field on the Efimov State

We successfully reproduced the derivation of Efimov energy levels in three-dimensional space. In subsequent discussions, we extended this derivation to Schwarzschild spacetime. By combining the Schrödinger equation in curved spacetime, we obtained the corrections to the Efimov energy spectrum under weak gravitational fields. Furthermore, using Riemann normal coordinate expansions, we rigorously derived the hyperradial equation in the presence of strong gravitational fields and applied first-order perturbation theory to compute the internal energy corrections of the Efimov system under strong-field conditions. We further propose a potential experimental approach to validate the theory by simulating spacetime curvature effects induced by gravity through the use of ultra-cold atoms in optical lattices under laboratory conditions.

gr-qc↗

arXiv2025

Clay Edges Are Dynamic Proton-conducting Networks Modulated by Structure and pH

Montmorillonite, a ubiquitous clay mineral, plays a vital role in geochemical and environmental processes due to its chemically complex edge surfaces. However, the molecular-scale acid-base reactivity of these interfaces remains poorly understood due to the limitations of both experimental resolution and conventional simulations. Here, we employ machine learning potentials with first-principles accuracy to perform nanosecond-scale molecular dynamics simulations of montmorillonite nanoparticles across a range of pH. Our results reveal clear amphoteric behavior: edge sites undergo protonation in acidic environments and deprotonation in basic conditions. Even at neutral pH, spontaneous and directional proton transfer events are common, proceeding via both direct and solvent-mediated pathways. These findings demonstrate that montmorillonite edges are not static arrays of hydroxyl groups but dynamic, proton-conducting networks whose reactivity is modulated by local structure and solution conditions. This work offers a molecular-level framework for understanding proton transport and buffering in clay-water systems, with broad implications for catalysis, ion exchange, and environmental remediation.

physics.chem-ph↗

arXiv2025

Hierarchical Learning-Enhanced MPC for Safe Crowd Navigation with Heterogeneous Constraints

In this paper, we propose a novel hierarchical framework for robot navigation in dynamic environments with heterogeneous constraints. Our approach leverages a graph neural network trained via reinforcement learning (RL) to efficiently estimate the robot's cost-to-go, formulated as local goal recommendations. A spatio-temporal path-searching module, which accounts for kinematic constraints, is then employed to generate a reference trajectory to facilitate solving the non-convex optimization problem used for explicit constraint enforcement. More importantly, we introduce an incremental action-masking mechanism and a privileged learning strategy, enabling end-to-end training of the proposed planner. Both simulation and real-world experiments demonstrate that the proposed method effectively addresses local planning in complex dynamic environments, achieving state-of-the-art (SOTA) performance. Compared with existing learning-optimization hybrid methods, our approach eliminates the dependency on high-fidelity simulation environments, offering significant advantages in computational efficiency and training scalability. The code will be released as open-source upon acceptance of the paper.

cs.RO↗