model predictive control for in-hand manipulation

本文提出了一种基于原语信息的采样模型预测控制框架，用于解决多指灵巧操作问题，通过使用低维操作原语和关节级残差优化提高采样效率。

Addressing the challenges of high-dimensional dynamic modeling and millisecond-level latency sensitivity in dexterous multi-objective manipulation, this paper proposes a Goal-Conditioned Probabilistic Model Predictive Control (GC-PMPC) framework. GC-PMPC integrates an asynchronous MPC architecture with an ensemble of probabilistic neural networks to explicitly model system uncertainty, while enabling joint optimization of tactile feedback and goal-conditioned policies. Evaluated on Shadow Hand simulation and the real-world DexHand 021 platform (12-DoF + 5 tactile channels), GC-PMPC achieves high-precision, real-time, adaptive manipulation of a cube toward arbitrary target poses—requiring only 80 minutes of training. It significantly outperforms state-of-the-art RL-based and deterministic MPC methods. The core contribution lies in the first deep integration of goal conditioning, probabilistic modeling, and asynchronous MPC—thereby simultaneously ensuring robustness, generalization across unseen goals, and sub-10-ms closed-loop response times.

To address the limited naturalness and responsiveness of tendon-driven prosthetic wrists, this paper proposes a computationally efficient model predictive control (MPC) framework enabling real-time, user-intent-driven adaptive regulation. The method innovatively integrates Euler–Bernoulli beam theory to formulate a kinematic model and employs Lagrangian mechanics to derive a dynamics model—balancing modeling fidelity with computational tractability for embedded implementation. Comprehensive simulation and experimental validation demonstrate substantial improvements in wrist joint trajectory tracking accuracy and dynamic response speed: operational error decreases by 32%, and subjective naturalness ratings increase by 41%. This work establishes an embeddable, lightweight, and highly adaptive control solution for intelligent prosthetic wrists, advancing practical deployment of responsive, user-centered upper-limb prostheses.

Modeling robot contact dynamics and achieving real-time control remain critical challenges in manipulation and locomotion tasks. This paper proposes a lightweight inverse-dynamics trajectory optimization framework that integrates contact-implicit modeling, efficient Hessian approximation, and sparse nonlinear programming structure, substantially reducing computational overhead for contact-sensitive model predictive control (MPC). The open-source real-time MPC solver achieves >100 Hz closed-loop control on a 20-degree-of-freedom bimanual robot platform—marking the first hardware demonstration enabling simultaneous high-dynamic legged locomotion and complex dexterous manipulation under robust contact control. Key contributions include: (i) overcoming the real-time feasibility bottleneck of contact-implicit MPC; and (ii) establishing a unified optimization paradigm that jointly ensures modeling fidelity, computational efficiency, and hardware deployability.

To address the generality and efficiency bottlenecks in robot motion planning and control under complex environments, this paper proposes a dynamics-free, sampling-based Model Predictive Control (MPC) framework. The method integrates the fully parallel GPU-accelerated physics simulator IsaacGym seamlessly into the Model Predictive Path Integral (MPPI) controller, enabling direct optimization of control sequences via forward-dynamics sampling. This implicit treatment of contact dynamics and multi-joint coupling eliminates the need for explicit dynamical modeling. Crucially, the approach supports plug-and-play transfer across diverse robot morphologies and object types. We validate its superiority on challenging tasks—including mobile navigation with obstacle avoidance, non-prehensile manipulation, and high-dimensional whole-body control—demonstrating both high computational efficiency and successful sim-to-real deployment. The implementation is open-sourced.

This work addresses the high computational cost of model predictive control (MPC) in real-time control of a 3-degree-of-freedom robotic arm by proposing a behavior cloning–based neural network surrogate. An expert controller is constructed by integrating inverse kinematics with MPC, and a static deep multilayer perceptron (MLP) is trained to emulate its input–output mapping. The study demonstrates that this static architecture outperforms temporal models, indicating that instantaneous state information suffices to replicate MPC behavior. Experimental results show a task success rate of 84.98% under relaxed tolerances, with inference latency reduced by a factor of three compared to MPC. However, under stringent accuracy requirements, the approach exhibits steady-state errors, revealing room for improvement in precision. These findings highlight the efficiency and potential of lightweight static networks for real-time robotic control.

该研究通过结合模型预测控制(MPC)与强化学习(RL)，解决了真实世界中灵巧操作任务早期探索效率低下的问题。

This work addresses the inherent conflict between high-precision trajectory tracking and compliant contact in dexterous hands under fixed-gain control. The authors propose a drive-agnostic impedance model predictive control (MPC) framework that algebraically reduces tendon-driven systems to a constant-coefficient double integrator via feedforward linearization. An encoder-only augmented Kalman disturbance estimator is embedded to eliminate steady-state errors. For the first time, a constant-\(A_d\) unbiased MPC architecture is applied to dexterous manipulation, achieving high-speed real-time optimization under hard constraints while ensuring stability and recursive feasibility with strong disturbance rejection. Experiments demonstrate that a hydraulic finger attains 0.5 mrad RMS tracking error and 0.1 mrad steady-state error, with only 6.6 mrad peak deflection under contact disturbances. On a 16-DOF LEAP hand, a 2.5 N grasping disturbance is rejected within 0.7 seconds, yielding performance improvements of tens to thousands of times over classical methods.

This work addresses the challenges of high training overhead and complex problem formulation in combining model predictive control (MPC) with reinforcement learning (RL) for humanoid robot motion control. To this end, the authors propose an MPC-RL framework that leverages a centroidal dynamics-based MPC during training to generate guiding trajectories and designs an MPC-informed reward mechanism to enhance learning efficiency. Furthermore, they develop πⁿMPC, a just-in-time compiled, batched GPU-parallel MPC solver capable of handling time-varying dynamics, which substantially reduces computational costs. Experimental results demonstrate that the proposed approach outperforms existing methods across a range of locomotion and manipulation tasks, achieving both high performance and computational efficiency on real hardware.

This work addresses the challenge of simultaneously maintaining contact and accurately tracking object contours in robotic contour-following tasks by proposing a vision-based tactile model predictive control framework (VBT-MPC). For the first time, model predictive control is directly applied in the contour feature space extracted from an eye-in-hand visuotactile sensor, eliminating the need for separate pose estimation or complex force-control modules. By integrating visuotactile perception with feature-based visual servoing, VBT-MPC achieves high-precision and stable contour tracking across objects with diverse geometries and material properties in both simulation and real-world experiments. This approach substantially simplifies the system architecture while significantly enhancing tracking performance.