arXiv subjects

Mingcong Li

Publications and source records attributed to Mingcong Li.

4 recordsLinked to original sources

arXiv2026

Spectrally Routed Diffusion Policy for Multi-Task Robotic Manipulation

Diffusion-based policies have achieved strong performance on single-task robotic manipulation. However, a single model trained jointly on multiple tasks consistently underperforms the per-task specialists trained on the same demonstrations. We focus on two challenges. First, the convolutional encoders commonly used in diffusion-based policies produce visual conditions that are not discriminative enough. Second, gradients from incompatible action modes update the shared denoiser's dense parameters along the same directions, causing destructive interference during joint training. To address both, we propose SR-DP (Spectrally Routed Diffusion Policy). A frozen VGGT encoder with a spatial-compression adapter produces compact multi-scale semantic-geometric tokens, a task-discriminative visual condition for tasks that share a workspace. We first train the denoiser as a standard diffusion policy, then decompose selected linear layers by SVD into a frozen basis of orthogonal rank-one experts, adapted through spectral routing and Householder input rotations. On six RoboTwin tasks, SR-DP averages a 68.4% success rate, surpassing multi-task VO-DP by 21.8 points and the strongest per-task baseline by 3.8 points. On four real-world bimanual tasks, Stage 2 reaches a 43.0% average success rate and improves over Stage 1 by 19.0 points.

cs.RO↗

arXiv2026

Gaze Prompts: Temporally Dense Human Attention for Vision-Language-Action Fine-Tuning

Vision-Language-Action (VLA) fine-tuning pairs images with actions at every step, yet typically provides only a task-level language instruction, leaving moment-to-moment visual relevance implicit. We introduce \emph{eye-tracker-supervised gaze prompting}, which uses gaze recorded during VR teleoperation to provide frame-level visual guidance for VLA fine-tuning. During training, recorded gaze locations are rendered as crosshairs on the robot's head-camera images. At deployment, a lightweight predictor estimates gaze locations from recent images and the instruction, supplying the same type of visual prompt without an eye tracker or changes to the policy architecture. Instantiated with $π_0$, gaze prompting increases mean success from $26.3\%$ to $56.0\%$ across six real-world bimanual manipulation tasks, with gains also observed when a single policy is trained on all six tasks. We release \textsc{GazeMani}, a dataset of $1{,}200$ teleoperated trajectories with synchronized gaze.

cs.RO↗

arXiv2025

A Risk-Aware Adaptive Robust MPC with Learned Uncertainty Quantification

Solving chance-constrained optimal control problems for systems subject to non-stationary uncertainties is a significant challenge.Conventional robust model predictive control (MPC) often yields excessive conservatism by relying on static worst-case assumptions, while standard stochastic MPC methods struggle when underlying uncertainty distributions are unknown a priori.This article presents a Risk-Aware Adaptive Robust MPC (RAAR-MPC) framework,a hierarchical architecture that systematically orchestrates a novel synthesis of proactive, learning-based risk assessment and reactive risk regulation. The framework employs a medium-frequency risk assessment engine, which leverages Gaussian process regression and active learning, to construct a tight, data-driven characterization of the prediction error set from operational data.Concurrently, a low-timescale outer loop implements a self-correcting update law for an adaptive safety margin to precisely regulate the empirical risk and compensate for unmodeled dynamics.This dual-timescale adaptation enables the system to rigorously satisfy chance constraints with a user-defined probability, while minimizing the conservatism inherent in traditional approaches.We formally establish that the interplay between these adaptive components guarantees recursive feasibility and ensures the closed-loop system satisfies the chance constraints up to a user-defined risk level with high probability.Numerical experiments on a benchmark DC-DC converter under non-stationary parametric uncertainties demonstrate that our framework precisely achieves the target risk level, resulting in a significantly lower average cost compared to state-of-the-art robust and stochastic MPC strategies.

eess.SY↗

arXiv2025

Online-Optimized Gated Radial Basis Function Neural Network-Based Adaptive Control

Real-time adaptive control of nonlinear systems with unknown dynamics and time-varying disturbances demands precise modeling and robust parameter adaptation. While existing neural network-based strategies struggle with computational inefficiency or inadequate temporal dependencies, this study proposes a hybrid control framework integrating a Temporal-Gated Radial Basis Function (TGRBF) network with a nonlinear robust controller. The TGRBF synergizes radial basis function neural networks (RBFNNs) and gated recurrent units (GRUs) through dynamic gating, enabling efficient offline system identification and online temporal modeling with minimal parameter overhead (14.5% increase vs. RBFNNs). During control execution, an event-triggered optimization mechanism activates momentum-explicit gradient descent to refine network parameters, leveraging historical data to suppress overfitting while maintaining real-time feasibility. Concurrently, the nonlinear controller adaptively tunes its gains via Jacobian-driven rules derived from the TGRBF model, ensuring rapid error convergence and disturbance rejection. Lyapunov-based analysis rigorously guarantees uniform ultimate boundedness of both tracking errors and adaptive parameters. Simulations on a nonlinear benchmark system demonstrate the framework's superiority: compared to PID and fixed-gain robust controllers, the proposed method reduces settling time by 14.2%, limits overshoot to 10%, and achieves 48.4% lower integral time-weighted absolute error under dynamic disturbances. By unifying data-driven adaptability with stability-guaranteed control, this work advances real-time performance in partially observable, time-varying industrial systems.

eess.SY↗