Autonomous navigation in unknown, complex indoor environments remains challenging due to limited sensing range and severe partial observability. Conventional methods rely on local maps without foresight, causing dead-ends and long detours, while local goal selection based on Euclidean distance or frontier coverage fails to balance efficiency with directionality. To address these challenges, we propose ForexNav, a foresight-enabled exploratory navigation framework. To handle structural ambiguity in unseen regions, we introduce Foresight Hypothesis Fusion (FHF), which maintains multiple WGANbased map predictions and reweights them via temporal evidence accumulation. A Traversability-aware A* search then quantifies predictive traversability on the fused map, enabling a multiobjective planner to synthesize path feasibility, kinodynamic conformity, monotonic-progress consistency, and geometric distance for optimal intermediate goal selection and dynamically consistent trajectory generation. Experiments in four simulated indoor scenes of up to 3,300 m$^{2}$ demonstrate navigation success while reducing total travel time by 25.0% and improving average velocity by 13.3% over the strongest baseline, with path ratio improvements of 22.2% on average in large-scale environments ($\geq$2,000 m$^{2}$). Real-world deployment on a quadruped robot supports practical feasibility, and extension to a fixed-altitude micro- UAV further suggests preliminary cross-platform transferability.
Learning from Demonstration (LfD) combined with Behavior Trees (BTs) aims to lower the programming burden required to construct robot task programs from demonstrations. However, existing approaches have two structural limitations: skills are bound to specific object instances with no mechanism for runtime rebinding, and inferred symbolic conditions are compiled as frozen constraints that cannot be independently verified or revised. This paper presents DEPT-BT, a framework that transforms single demonstrations into parameterized, executable, and maintainable behavior tree programs. DEPTBT introduces anchor-based parameterized skill operators that decouple skills from demonstration objects, enabling task-level generalization through runtime parameter rebinding. The framework also incorporates a verify-revise closed-loop mechanism that maintains symbolic conditions as editable entities, supporting targeted failure attribution and correction without complete redemonstration. Experiments on four manipulation tasks yield an overall success rate of approximately 90%. DEPT-BT attains 95% success when transferring skills to novel objects, compared to 0% for the baseline, and the revision mechanism reduces failure recovery time by 64% compared to full re-demonstration.
This study proposes a motion trajectory correction method for a robot based on iterative learning and force information to automate contact-based glue application tasks. To apply glue using a sponge, force information is essential because a uniform coating requires subtle changes in the contact force at the fingertips. Most studies on the automation of coating tasks focus on noncontact processes such as spray painting that do not require force sensing at the end effector. By contrast, we aim to automate contact-based glue application tasks using a robot-teaching system that matches the forces produced by a human operator and robot system. In a similar system, force in the pressing direction of the sponge was successfully reproduced. However, force errors remained along the other axes. Hence, in our system, the target motion trajectory is also corrected to match the torques about the two axes orthogonal to the pressing direction, improving the accuracy of the robot's movements. Experimental results demonstrate the effectiveness of the proposed method and its potential for use in contact-based coating tasks. In addition to the force and torque evaluation, a water-droplet spreading task on color-changing paper is used as a task-level surrogate assessment; applying the correction improved the spreading uniformity relative to no compensation.
Accurate human trajectory prediction is essential for autonomous driving and robot navigation. Although deep learning-based approaches have achieved substantial progress, they often suffer from the train–inference gap, which arises from distributional discrepancies between training and test environments. The goal-guided framework has recently emerged as an effective paradigm to mitigate this issue by exploiting the scene-invariant prior that human motion is typically goal-driven. However, existing goal-guided approaches do not fully leverage additional scene-invariant priors and therefore still face two key limitations: insufficient diversity in predicted goals and the lack of explicit modeling of human kinematic consistency. In this paper, we propose SIPTraj, a goal-guided trajectory prediction framework that further integrates scene-invariant priors at both the long-term intention and short-term motion levels. For goal estimation, we introduce a set of goal candidates derived from offline clustering of trajectory endpoints as a data-driven approximation of scene-invariant long-term intention modalities, and develop a weighted Farthest Point Sampling (FPS) strategy to select these candidates by jointly considering spatial coverage and predicted likelihoods, thereby improving goal diversity while preserving semantic plausibility. For trajectory completion, we develop a kinematic-aware dual-stream autoregressive decoder that jointly predicts aligned position and velocity in two streams. The decoder adopts a kinematic initialization with residual refinement (KIRR) strategy: future states are first initialized using a constant-velocity kinematic model and then refined via reciprocal cross-attention, which injects kinematic consistency into the decoding process and leads to more natural and accurate trajectory prediction. Extensive experiments demonstrate that our approach achieves state-of-the-art performance on the ETH/UCY and SDD benchmarks.
Accurate trajectory tracking is a fundamental control task in robotics, typically achieved using model-based methods. Computed Torque Control (CTC) is computationally efficient but struggles to handle system constraints, while Model Predictive Control (MPC) can manage constraints but often results in nonlinear formulations. This work introduces a convex Model Predictive Control (MPC) formulation that incorporates learned dynamics for a robot manipulator. A nonparametric model is used to learn the residual between a nominal model and the observed dynamics, using an inverse dynamics formulation. Convexification is achieved by applying implicit partial differentiation to linearize the inverse dynamics model augmented with the residual model along the reference trajectory, resulting in a Linear Time-Varying (LTV) system. Extensive simulations and hardware experiments on a Franka Emika Panda manipulator demonstrate that our convex MPC approach outperforms learning-based baselines in handling a broad range of model mismatches and constraints. Results also show that learning the residual in inverse dynamics form offers better tracking performance than learning the residual in forward dynamics.
In manufacturing and construction industries, heavy objects lifted with cranes are manually positioned by floor workers pushing and pulling the object to bring it to a desired location. Skills are required, especially for mating a crane-suspended object with a fixture, a process called fixture loading. This paper aims to automate this dangerous and difficult task using a multi-cable crane capable of contact-rich dexterous manipulation, the DexCrane. A prototype DexCrane suspends an object and guides it to a target pose through a series of contacts with jigs. This is accomplished by taking advantage of manipulation with cables. When jigs are touched, the cables become slack and thus exhibit a compliant behavior without active force and compliance control. A trajectory planning method is presented to generate trajectories so that a cable-suspended object can be guided to a target pose through a series of jig contacts. Accessibility conditions are analyzed to determine whether the cable-suspended object can reach the target pose while interacting with the jigs. Dynamic Programming (DP) is applied to trajectory optimization subject to contact constraints and cable slack conditions. The proposed approach is implemented on a DexCrane prototype and validated through hardware experiments and simulations, demonstrating its effectiveness and practical feasibility.
While safety in robotics has historically been analogous to avoiding contact, robots performing high-speed manipulation tasks may be required to make contact with the environment or themselves at speed. We propose a method to allow but constrain impact using an impact-limiting hybrid control barrier function (IHCBF). Whenever a collision is imminent, the IHCBF implicitly adapts both velocity and directional manipulator inertia to reduce impact impulse while explicitly regulating momentum. The CBF framework naturally behaves as a safety filter, constraining impact impulse under any choice of task-specific control. We validate the proposed IHCBF through theoretical analysis and demonstrate its minimally invasive safety via simulations and hardware experiments of controlled collisions. Our framework provides a path toward dynamic robot behaviors with applications to high performance and safe manipulation.
AI planners can now generate a task plan for a team of robots, yet a human operator must stay in control of what the robots do. Doing so requires an interface that makes the plan visible and each step actionable, so mistakes can be caught before they reach the robots. Today, such an interface is developed for each specific application. This paper shows that non-expert operators can supervise an AI-generated multi-robot plan, authorizing sound plans and correcting faulty ones before any robot acts. The interface is generated automatically by HRI grounding, an architecture that renders any plan in a pre-defined format as one control per step, each showing in plain language what a robot will do to which target and holding the command it runs on the operator's confirmation. Across 100 generated plans over two domains, five planner models, and two interface engines, HRI grounding produced a working interface and dispatchable ROS commands for nearly every plan, with no specific interface developed for any particular plan, domain, or planner. About one in five plans contained a semantic fault. In a study with sixteen operators, participants distinguished correct plans from faulty ones with high sensitivity and few false alarms ($d^{\prime }\!=\!2.90$), and through the interface recovered 85 of the 96 missions that would otherwise have failed unsupervised. Operating a mission through the HRI grounding interface also lowered completion time, interactions, errors, and workload relative to a manual baseline. Commands dispatched by HRI grounding also drive a physical Unitree Go2 under ROS with live plan regeneration.
Mobile robots often face challenges operating in partially known environments where success depends on factors such as robot capabilities, task characteristics, and the environment. This letter presents a generalized, modular framework that enables robots to explicitly account for these factors through the concept of a mission-gain: a pose dependent metric quantifying mission performance (e.g. visibility). When the environment can be proactively explored, a cost map encoding the value of the mission-gain is generated and integrated into a cost function for a model predictive path integral (MPPI) controller, which supports arbitrary mission-specific cost functions. The controller generates inputs that balance progress toward the goal with improved mission performance. The map generation is decoupled from the control execution, making the framework modular and well-suited for heterogeneous multi-robot systems. To address transmission demands of large maps, we distill the map into a Gaussian mixture model (GMM) approximation, yielding a representation with far fewer parameters. The approach is validated in simulations and lab experiments across two scenarios: occlusion-aware navigation and vibration minimization over rough-terrain. The results for both scenarios show the mission-gain aware robot successfully reaching the goal while improving mission safety.
Robotic vascular cannulation is inherently uncertain: both the needle and vessel are compliant, vessels can translate or roll under contact, and small misalignments can cause surface slip, tenting, or double puncture rather than intraluminal entry. This paper presents a recovery-oriented robotic cannulation framework that couples multimodal sensing with a bi-rotational insertion strategy. A robotic platform autonomously segments the vessel centerline from microscope images, estimates vessel surface and lumen geometry from OCT B-scans, and executes a staged alignment and insertion trajectory. During insertion, pressure sensing provides access confirmation, while real-time vessel tracking and an insertion threshold identify deformation-driven failures that terminate the current attempt and prompt a user-confirmed reset/retry sequence. We evaluate the framework on ex vivo perfused chicken thigh vessels under matched bi-rotational and straight insertion conditions. Bi-rotational insertion increases cannulation success from 76% with straight insertion to 96%, and reduces the incidence of higher-risk failures such as slip-related misalignment and back-wall perforation observed in straight insertion. These results highlight the value of combining state-aware recovery logic with motion-level actuation choices to improve robustness of autonomous vascular access in soft tissues.
We introduce a novel geometric interpretation of camera-radar online alignment in dynamic maritime environments by formulating it as conic-section-constrained scalar field correlation using visual boundary cues. The key novelty lies in reinterpreting camera boundary cues and radar observations through the cone-plane intersection geometry between the camera's projection cone and the radar's slicing plane. This geometry yields an analytic mapping from camera-derived boundary cues to conic curves on the radar plane. We then maximize a scalar field correlation objective to estimate alignment parameters for each frame. In this work, online alignment refers to the frame- wise estimation of roll, pitch, and sensor height with respect to the radar-referenced sensing plane. Experiments on the Pohang Canal Dataset show that CROCS achieves the best overall radar-referenced projection alignment among the evaluated methods across RGB and IR sequences, while maintaining runtime compatible with the radar scan period on the evaluated platform.
Accurately modeling the mechanics of concentric tube robots (CTRs) remains challenging due to nonlinear curvature coupling, torsional interactions, and sensitivity to material and geometric parameters. Physics-informed neural networks (PINNs) offer a flexible alternative by embedding governing equations directly into the training process. In this work, we develop a PINN for modeling the free-tip kinematics of a three-tube CTR with a max length of 250 mm, incorporating geometric compatibility, constitutive relations, and equilibrium equations based on Cosserat rod theory into the training pipeline. We additionally introduce a Fourier-based joint encoding to represent the actuation rotations. Along 20 equally spaced points for each backbone configuration, the PINN achieved mean and maximum errors of 0.80$\pm$ 0.42mm and 1.71$\pm$1.12 mm, outperforming the analytical Cosserat model and demonstrating that physics-informed learning can approximate nonlinear CTR mechanics. A dataset-size ablation against a standard neural network showed that the two networks perform comparably when the training data densely cover the workspace, whereas the PINN was substantially more accurate under sparse coverage, retaining error at or below the analytical model with as little as 5% of the data.
Data-driven bimanual dexterous grasping remains challenging due to the lack of large-scale physically valid datasets and the difficulty of generating feasible bimanual grasps from partial observations. We propose an efficient bimanual grasp synthesis pipeline that decouples bimanual grasp generation into single-hand grasp optimization and Cartesian pairing, followed by geometric filtering and physics-based validation. Using this pipeline, we construct BiDexGen-Dataset, a large-scale synthesized bimanual grasp dataset containing millions of physically valid bimanual grasps across 2,397 objects. To exploit the synthesized data for real-world deployment, we further introduce BiDexGen, a diffusion-based generative model that directly generates feasible bimanual grasp configurations from partial object point clouds. To improve grasp generation from real-world partial perception, we design a contact-aware sampling strategy for training data construction that collects simulated observation-grasp pairs by aligning partially visible object regions with corresponding feasible grasps, enabling BiDexGen to generate feasible bimanual grasps from partial observations. Experiments show that our synthesis pipeline outperforms prior bimanual grasp synthesis methods in simulation success rate and generation efficiency. Trained on BiDexGen-Dataset, BiDexGen achieves 68.47% success in simulation and 71.00% on a real dual-arm robotic system.
This paper proposes a novel in-hand rotational manipulation method using a multi-fingered hand equipped with a cloth at the fingertips, applicable to a wide range of objects without object-specific modeling or planning. By leveraging the cloth's local adaptability to object shapes and its global surface shape to determine object motion direction, the method enables vertical-axis rotation of objects with diverse shapes. First, we introduce an analytical model and define a cloth-configured energy landscape (CCEL) as a unified energy-based representation of feasible cloth surface shapes under constraints imposed by fingertip positions. Next, we compute CCELs for various fingertip positions and construct CCEL-based rotation modes. In these modes, the gradients of selected CCELs guide the direction of object motion, while fingertip vibrations induce object motion. Finally, we develop a four-fingered cloth-attached hand and conduct CCEL generation and rotation experiments. We demonstrate that rotation of objects with various shapes is achieved by selecting appropriate modes without object-specific modeling or planning, highlighting the versatility of the proposed method.
Kinematic calibration is essential for improving the absolute positioning accuracy of industrial manipulators. Although Product of Exponentials (POE)-based methods have become the academic standard because of their singularity-free representation of joint axes, their adoption in industry remains limited. Most commercial controllers are architected around Denavit-Hartenberg (DH) parameters, so projecting identified 6-degree-of-freedom (DoF) screw errors onto the constrained 4-DoF DH model after optimization is an ill-posed operation. Conventional solvers are also prone to overfitting the stochastic measurement noise produced by low-cost industrial sensors. This paper addresses both issues with a framework termed DH-LPOE. Using the dyad kinematics of the Local POE (LPOE) formulation, an analytical Jacobian is derived that embeds the DH constraints directly in the identification loop, and the measurement noise is parameterized as an independent decision variable within a regularized objective function. The resulting corrections are therefore physically realizable and directly compatible with commercial controllers, while parameter overfitting is substantially reduced. Simulations and physical experiments on a serial manipulator under macroscopic measurement noise show that DH-LPOE improves robustness and data efficiency, reducing workspace error with a compact dataset of only 20 postures.
Robust implicit flocking for nonholonomic swarms using noisy position-only observations remains challenging because noise suppression can delay maneuver-intent tracking. Existing methods often rely on instantaneous kinematic variations, making them noise-sensitive or introducing alignment delays through state estimation. We propose Geometric Alignment for Intent in Nonholonomic Swarms (GAIN), a framework that infers steering intent from the swept area of relative-position vectors, a temporally integrated geometric representation that extracts intent without explicit higher-order state estimation. GAIN combines this representation with a predictive feedforward alignment mechanism to mitigate kinematic phase lag and a truncated repulsive field to reduce noise-induced false interactions. Simulations and experiments with a 30-robot physical swarm show that GAIN enables rapid propagation of evasive maneuvers from a single anonymous leader while improving responsiveness, noise resilience, and formation cohesion under noisy, communication-constrained conditions and bounded turning-rate constraints using local neighbor observations.
Pneumatic soft robots have found widespread applications across various fields. Recently, researchers have explored integrating actuation, sensing, and computation directly into pneumatic logic systems, thereby eliminating the need for bulky electronics. This approach makes the systems more compact and better suited for operation in humid or harsh environments where conventional electronics may fail. Most electronics-free soft robots were built using pneumatic ring oscillator circuits, in which the output of each unit sequentially triggers the next. Continuous oscillation of ring oscillator requires an odd number of units (2n + 1, n $\geq$1), where the final unit feeds back to the first to sustain oscillation. This constraint limits design flexibility and increases system size, as a minimum of three units are required to form a functional ring. Here, we propose an innovative electronics-free pneumatic oscillator capable of operating with an arbitrary number of units. Specifically, we demonstrate self-excited oscillation in a singleunit configuration, mutual oscillation between two coupled units, and multi-unit oscillations (3, 5, 10 units) in ring architectures. This design fundamentally eliminates the 2n + 1 constraint of conventional ring oscillators, significantly expanding the freedom of electronics-free pneumatic systems and offering a new pathway for embodied and adaptive control in soft robotics.
Reinforcement learning shows strong potential for bipedal locomotion control, while reduced-order models provide compact and interpretable physical priors. This paper presents an Angular Momentum Linear Inverted Pendulum (ALIP) guided reinforcement learning framework for dynamic locomotion of a small point-foot bipedal robot, where foot placement references generated online by ALIP serve as structured physical guidance for policy training. Compared with end-to-end learning without reference guidance and a counterpart guided by the conventional linear inverted pendulum based model under matched training settings, ALIP guidance shows improved center-ofmass height regulation and enhanced robustness to payload and unseen-terrain disturbances. Furthermore, a speed-conditioned step duration modulation mechanism constrained by ALIP gait parameters and hardware-feasibility is introduced to raise the maximum achievable speed and improve speed tracking accuracy. The framework is validated in both simulation and hardware experiments, where the robot achieves a maximum forward speed of 2.26m/s (Froude number 1.30) and stable locomotion with a payload of 37.5% of the nominal robot mass, with additional qualitative demonstrations of stable traversal across diverse unseen terrains.
Human intent recognition is crucial for effective robotic teleoperation, yet existing methods rely on a predefined set of intentions, a requirement that is manually intractable in complex open-world scenarios. We propose an unsupervised framework that integrates Bayesian nonparametric learning with multi-modal Large Language Models (LLMs) to autonomously discover and interpret human intentions. We first employ a hierarchical Dirichlet process hidden semi-Markov model (HDP-HSMM) to adaptively segment kinematic sequences into motion primitives, reducing the load on downstream models. Each segment is represented by multi-modal features extracted from front-facing video, robot kinematics, and eye fixations. We then develop a reasoning architecture based on zero-shot Chain-of-Thought (CoT) prompting, where a Vision-Language Model (VLM) translates each segment into a semantic description, and an LLM clusters these descriptions to reveal intentions and generate human-readable interpretations. Experimental results on real-world data show that our method can uncover meaningful intentions highly consistent with expert judgments, outperforming baselines in intra-intent consistency by achieving the lowest semantic outlier rate of 8.5%.
Lightweight micro aerial robots are increasingly used in confined indoor environments, yet perception of nearby obstacles remains difficult under strict payload constraints. Inspired by human perception of sound changes near surfaces, we propose a lightweight passive-acoustic pipeline that detects nearby reflectors directly from quadrotor ego-noise. The method uses a lightweight detector with consistency gates to infer nearfield obstacle cues without a reference spectrum, data-driven training, or environment-specific calibration. We validate the pipeline with COMSOL-based simulations and real-flight evaluations across different quadrotor configurations, obstacle materials, ultrasonic and state-of-the-art acoustic/learning baselines, and closed-loop avoidance tests. The test results indicate partial generalization across platforms and highlight the potential of egonoise- based passive acoustics for low-cost perception of nearby obstacles.
