When GNSS returns after an outage, the estimator faces an ambiguous discrepancy: the gap between the new fix and the LiDAR-inertial odometry (LIO) can be drift that LIO accumulated while GNSS was away, or a multipath-biased GNSS return. A residual-only layer cannot tell the two apart, and either mistake is costly: rejecting a valid fix strands the filter on drifting odometry, accepting a biased one corrupts the state. We propose AMML (Adaptive Measurement Model Layer), a lightweight rule for the widespread NavSatFix-only setting in which the integrator receives only a fused position fix and its covariance. AMML budgets how much drift LIO can plausibly have accrued over the denial (duration $T$, travelled length $S$), evaluates the observed discrepancy under a drift-inflated Rayleigh $H_{0}$ against a bias-slab Rice $H_{1}$, and maps the sequential posterior to a single GNSS covariance scale; the platform drift coefficient is fitted online. Across a 1650-run synthetic fault panel, seven real-bag experiments on four S3E sequences, a KITTI denial sweep, and a cargo-robot sequence with naturally occurring multipath and RTK loss, AMML matches residual-only baselines when GNSS is clean and avoids their failure modes when the denial history matters, e.g. $152\times$ lower post-return RMSE than a memoryless NIS gate after a long denial.
A central challenge in bio-inspired swarm control is to transform biological group-motion trajectories into local rules that artificial agents can execute online. Existing studies have either prescribed interaction laws or inferred statistical associations from animal trajectories, leaving a gap between asymmetric predictive relations in real motion data and deployable swarm-control rules. This paper presents a trajectory-driven framework for learning executable local interaction rules from bird-flocking trajectories. Signed planar lateral acceleration is used as the maneuver variable, and surrogate-tested conditional transfer entropy extracts evidence of directed predictive dependence from perception-constrained focal–neighbor pairs. The retained evidence is distilled into a dual-output local kernel that is queried online to gate and weight visible neighbors. Paired closed-loop simulations show better continuous-maneuver tracking than adaptive-alignment and inertial-turning baselines. They also show that a cue applied to one agent can spread through the group. We then use a six-UAV platform to compare T2R-Swarm with WiSt under local light cues. T2R-Swarm propagates the induced turns faster while maintaining comparable polarization. These results provide a trajectory-to-rule pathway from directed predictive dependence in real group motion to executable local rules for swarm robotics.
Robotic surfaces composed of discrete actuator arrays offer promising platforms for distributed object manipulation, yet achieving robust transportation remains challenging due to localized contacts and actuator discreteness. We present a robust method for object transportation on discrete pin-array robotic surfaces, leveraging a discrete surface gradient-based control strategy. The proposed approach explicitly models the interaction between surface deformations and object motion, and computes a discrete gradient that accounts for localized height variations when the object size is comparable to the actuator size. In addition, three basic surface shapes-pit, bump, and ramp-are evaluated for their effectiveness in manipulating diverse objects. To improve robustness under uncertainties, an active disturbance rejection control (ADRC) framework regulates surface amplitude. Systematic experiments on a 6×6 robotic surface demonstrate successful point-to-point transportation, geometric path following, and reliable manipulation of both rigid and soft objects. Among the tested shapes, the pit consistently provides the highest transport stability and accuracy. The combined approach achieves high success rates across various object geometries and substantially improves performance for irregular objects compared to prior approaches, while it reaches as low as a $60\%$ success rate and poor path following performance for slender objects such as rods.
Model Predictive Control (MPC) has been widely employed in mobile robotics to compute optimal control inputs based on predicted future states. However, inevitable disturbances in real-world environments cause model mismatches that degrade the predictive performance of MPC. To address this issue, this paper proposes an On-Manifold Adaptive Law (OMAL) to estimate unknown disturbances and mitigate the effects of the resulting model mismatches. OMAL estimates physically consistent disturbances by explicitly respecting the compound manifold structure of the state space, thereby avoiding attitude-representation singularities through exponential and logarithmic maps. The disturbances estimated by OMAL are incorporated into the nominal model to improve prediction accuracy over the prediction horizon, while preserving the quadratic program (QP) structure of MPC and hence its feasibility under constraints. Simulations and real-world experiments with quadrotors demonstrate that the proposed OMAL reduces the effects of disturbances without prior disturbance models or additional sensors, thereby improving tracking performance during agile trajectories that involve aggressive maneuvers.
Existing drone landing systems, whether model-based or learning-based, typically rely on predefined visual markers (e.g., ArUco tags) or are trained for specific platform categories, and often assume structured target motion. Recent open-vocabulary studies enable text-prompted target selection, but integrating this capability with learned closed-loop landing control remains underexplored. To this end, we present Ask to Land, a mask-conditioned framework that enables a quadrotor to land on various platforms described by a user in natural language, including randomly moving platforms. The core of our approach is the integration of a frozen state-of-the-art open-vocabulary perception frontend (YOLO-World with MobileSAM) and a control policy learned by differentiable physics. By providing real-time, task-agnostic semantic segmentation masks as policy inputs, our framework enables the drone to land on prompt-selected targets without target-specific fine-tuning while continuously tracking dynamic target motion during descent. The frozen frontend processes the prompt, while the policy receives the mask without text embeddings or category labels. To efficiently optimize the control policy, we leverage differentiable physics learning via Backpropagation-Through-Time (BPTT), and employ a carefully designed reward that enforces visual alignment with the landing target and altitude-adaptive descent speed, together with stable control. Real-world deployments validate that our approach achieves landings on diverse user-specified objects, including static and moving platforms.
