Guaranteeing the safety of autonomous robotic systems under strict actuation limits and external disturbances remains a fundamental challenge in modern control design. In this work, we propose a novel learning framework to synthesize formally verified Robust Control Barrier Functions (RCBFs) that explicitly certify safety and control feasibility under worst-case disturbances. We employ smooth fully connected networks to maintain control continuity and introduce a Taylor expansionbased verification pipeline. Furthermore, we implement an automated binary search to precisely quantify the maximum disturbance the synthesized RCBF can withstand. Through simulations on a double integrator and 2D/3D quadrotors, our method verifies the RCBFs faster than off-the-shelf neural network verification solvers. Finally, we validate its practical applicability by achieving real-time, collision-free LiDAR navigation on a Unitree Go2 quadruped robot.
Accurately estimating robot state, i.e., pose and joint angles, from a single image remains challenging for in-the-wild scenarios, where 3D state labels, calibrated sensor rigs, and joint encoders are often unavailable in practice. Prior learning-based methods rely on synthetic supervision and sparse 2D cues from real images, i.e., keypoints and masks, which provide limited constraints for accurate monocular robot state estimation. To mitigate this, we present DRIP, a novel robot-fitting in the loop training framework that learns monocular robot state estimation from real images without 3D state labels. DRIP seamlessly integrates a feed-forward network with iterative robot state refinement in the training pipeline: the network provides an initial robot state prediction, which is then refined through an optimization stage that leverages dense visual cues, i.e., structure, geometry, and silhouette. The refined states are then used as pseudo-labels for subsequent training, forming a self-improving loop that transfers refinement benefits to the feed-forward model while preserving efficient single-shot inference at test time. Across multiple real-world robot datasets, DRIP outperforms prior approaches that rely on sparse self-supervision. We show that robot-fitting in the loop training enables effective real-image adaptation for monocular robot state estimation by generating reliable state-level pseudo-labels for stronger supervision.
Autonomous robots require real-time, multi-task scene understanding from raw sensor data, yet existing models demand large annotated datasets and full retraining whenever the deployment environment changes. We introduce a new few-shot multi-task scene understanding setting in which a single model is adapted to a new environment from a handful of annotated images, and propose SAGE, a few-shot adaptive model designed for this setting. SAGE jointly predicts semantic segmentation, depth, and surface normals from a single RGB image through a shared query-based architecture with task-specific decoding heads and the segmentation head classifying queries against CLIP text embeddings of the target classes. The model is first pre-trained on Hypersim, a large-scale photorealistic synthetic dataset, to acquire a general semantic-geometric representation, and then adapted to a target environment efficiently via LoRA, updating only a small subset of weights while preserving the jointly learned representations. Experiments on NYU Depth v2, Cityscapes, DIODE, KITTI and VKITTI2 show that SAGE adapts across all three tasks with as few as 24 annotated images, matching or surpassing state-of-the-art methods trained with substantially more data.
The integration of 3D tactile sensing into dexterous robotic fingertips remains an open challenge due to stringent constraints on confined volume, curved surface, and fixed mechanical interfacing. This work introduces μ-Tac, a novel fingertip-scale vision-based tactile sensor engineered as an ultra-compact retrofit module, preserving the biomimetic fingertip shape at the diameter less than 9 mm. An automated traversal calibration protocol was proposed, facilitating autonomous acquisition of high-quality “Optical Flow-Force” data across the entire curved sensing surface. A Pyramid Pooling-based multi scale force regression model was then developed, to directly map dense optical flow fields to the resultant 3D contact forces, encompassing both normal and shear components. The accuracy of 3D force estimation for Fx, Fy, and Fz on held-out contact poses showed R2 values of 0.988, 0.993, and 0.984, respectively. Real world deployment experiments utilizing the force estimation model in complex scenarios, specifically fabric texture discrimination and deformable balloon grasping tasks, demonstrated a classification accuracy exceeding 90% across 40 fabric types and real-time tactile feedback in closed-loop control adjustments, with a mean disturbance-to-joint response latency of 74.25 ms. These results demonstrate significant potential of miniaturized vision-based 3D force sensing technology towards dexterous manipulations with tactile robotic hand.
This paper presents a novel lightweight variable stiffness mechanism, termed “String Binder,” which binds multiple pairs of elastic wires with a thin string. Pneumatic pressurization mechanically couples the wires, increasing the second moment of area of wires and bending stiffness. Using a cross-sectional model, the second moment of area of the wires was evaluated for the two stiffness modes. Seven prototypes with different wire diameters were fabricated and evaluated using a constant-moment bending test, achieving up to 18.2 times stiffness increase with a mass below 45 g. Experiments agreed well with the model within a certain range, and a specific stiffness demonstrated superior stiffness efficiency over conventional mechanisms. These results demonstrate the potential of the String Binder for compact continuum robotic and minimally invasive applications.
