active haptic exploration

This work addresses the challenge of inaccurate object recognition and six-degree-of-freedom pose estimation in tactile interaction, which arises from perceptual ambiguity and contact uncertainty. To this end, the authors propose a Bayesian active tactile exploration framework that unifies multimodal tactile signals—specifically GelSight images and wrist force/torque measurements—with non-contact free-space constraints within a joint belief update mechanism. The framework employs a tailored particle filter and an information-maximizing exploration strategy subject to reachability constraints to dynamically optimize subsequent touch actions. Experiments on a Franka Panda platform with 11 YCB objects demonstrate that the proposed approach significantly improves accuracy and robustness in both recognition and pose estimation compared to baseline methods relying solely on force/torque sensing, while also reducing the number of required interactions.

This work addresses the challenge in robotic manipulation where object shape and pose are often unknown due to sensor noise and occlusion, rendering conventional planning methods ineffective. The authors propose a unified framework that integrates tactile perception, geometric modeling, and motion planning: active tactile exploration is guided by Bayesian optimization to approximate object boundaries online using superellipses; an adaptive manipulation potential field encodes geometric constraints; and object pose is inferred in real time through an ordinary differential equation (ODE)-based estimator. Evaluated on 2D sorting tasks, the approach demonstrates strong robustness and generalization across diverse object geometries, achieving efficient real-time shape estimation and pose inference in both simulation and physical multi-arm robotic platforms.

This paper addresses the challenge of online estimating the pose and shape of unknown rigid objects using only tactile sensing during robotic manipulation. We propose an active tactile exploration framework that integrates physics-constrained modeling—thereby avoiding numerical stiffness from rigid contact—and an information-gain-driven exploration strategy, leveraging differentiable physics simulation and gradient-based optimization for joint pose and geometric identification. Our key contributions are: (i) a unified objective function jointly optimizing physics constraint violation loss and expected information gain; and (ii) a low-mobility-cost hybrid sampling scheme combining stochastic and active strategies. Experiments on both simulation and real robotic platforms demonstrate that high-fidelity reconstruction of cube and convex polyhedral objects—achieving accurate pose and shape estimation—is accomplished within approximately 10 seconds of tactile interaction, significantly improving online recognition efficiency and robustness.

This work addresses the challenge of accurate object pose estimation in contact-rich manipulation tasks, where purely vision-based methods often fall short and existing vision-tactile fusion approaches suffer from reliance on costly offline training and poor generalization. The paper proposes BayesContact, a novel framework that introduces simulation-based inference into vision-tactile pose estimation for the first time. It employs particle filtering to maintain an online belief over object pose, fusing depth images with contact evidence derived from force/torque measurements. Crucially, a physics simulation-driven forward model computes observation likelihoods, enabling active probing and pose refinement without retraining. Experiments demonstrate that BayesContact significantly improves pose observability and task success rates—by approximately 30%—over vision-only baselines in both simulated and real-world robotic peg-in-hole tasks.

This work addresses object-agnostic 3D shape reconstruction from sparse tactile observations. Method: We propose a lightweight reconstruction framework integrating a single-view coarse visual prior with Bayesian active tactile exploration. First, we design an object-agnostic Bayesian active tactile policy that jointly optimizes information gain and contact failure avoidance under minimal contact budget. Second, we introduce a two-stage deformable mesh fitting pipeline: geometric initialization from a single view, followed by uncertainty-aware mesh optimization to ensure global structural consistency while capturing local deformations. Contribution/Results: Evaluated in simulation and on real robotic platforms, our method reduces required contacts by 37% and decreases reconstruction error in deformable regions by 52% compared to baselines. It significantly improves accuracy and robustness under sparse tactile data, enabling high-fidelity geometric perception for dexterous manipulation.

本文针对仅触觉条件下6自由度物体位姿精化问题，提出了一种基于物理信息的滑动窗口粒子滤波方法，通过融合最近触觉帧和考虑对称性来提高位姿估计精度。

This work addresses the challenge of building vision–tactile world models capable of supporting contact-intensive manipulation tasks for stable, long-horizon robotic planning. To this end, the authors introduce ContactWorld, a benchmark comprising twelve tasks designed to systematically evaluate the performance of diverse world models. Their analysis reveals that spatially structured and temporally coherent representations are critical for effective planning, and that tactile benefits hinge on cross-modal representational compatibility rather than merely increasing modality scale. Experimental results demonstrate that incorporating point cloud observations boosts average planning success rates from 20.7%–22.0% to 32.1%, and further integration of tactile force field representations elevates performance to 36.1%, substantially outperforming existing approaches.

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.

This study addresses the low sample efficiency and contact-dynamics-deviating exploration inherent in reinforcement learning for robotic manipulation by proposing TacEx. This framework introduces a novel tactile-channel-based epistemic uncertainty decomposition mechanism that anchors intrinsic curiosity to tactile feedback. By integrating multimodal uncertainty quantification with vision-language-action model post-training techniques, TacEx guides agents toward efficient exploration of complex contact dynamics. Experimental results demonstrate that TacEx achieves unsupervised, highly efficient grasping while significantly improving both the sample efficiency and generalization performance of downstream manipulation policies.

This work addresses the challenge of stable dexterous manipulation under conditions lacking external sensing, where contact uncertainties and gravitational disturbances often compromise grasp stability. The authors propose a novel approach that synergistically integrates reinforcement learning with mechanical design: a global grasp quality prior derived from classical grasp analysis is incorporated into the reward function, while fingertip curvature geometry is engineered to optimize local contact interfaces. This co-design strategy uniquely embeds two complementary physical priors—global grasp stability and local contact mechanics—into both the learning framework and hardware morphology. Experimental results demonstrate significant improvements in rotational efficiency, grasp robustness, and disturbance rejection across four palm orientations and three object types, with enhanced sim-to-real transferability for rolling manipulation tasks.