NASA's Curiosity rover uses cameras to provide scientists with ground-level views of Mars. The rover has investigated Gale Crater, finding rock layers that suggest past environments capable of supporting life. Scientists are now analyzing imagery to understand geological features like the polygonal fractures on the Martian surface. The process involves complex planning, engineering simulations, and sending numerous commands to the rover. This work helps advance the understanding of ancient habitability on Mars.
WM-VS introduces a target-centric world-model framework to address the prediction-control mismatch in closed-loop visual servoing. It uses offline correspondences to define a four-dimensional servo coordinate for translation, scale, and rotation. The method trains a reactive policy using action imitation and consequence supervision to favor error contraction. Experimental results show significant error reductions when closing the loop with the learned progress signal. This approach effectively links progress-aligned action consequences to repeated closed-loop correction.
AeRove is a compact bimodal aerial-terrestrial robot designed for close-range pipeline inspection, addressing limitations of standard drones. It achieves rapid reconfiguration between ground and flight modes in 200 ms using a spring-loaded bistable mechanism. Propeller guards function as wheels, enabling ground rolling for traversal, which reduces current draw by 14 times compared to continuous flight. Autonomous trials demonstrated navigation and leak detection on steel pipes, showing superior gas detection performance in both perched and aerial conditions. The system leverages propeller-induced airflow for enhanced gas sensing during flight.
Interactions, like shaking, can reveal hidden physical properties of objects inaccessible through conventional sensing. This study uses interaction-induced dynamic responses to infer object physics relevant to robotic manipulation. A soft robotic arm acts as a physical reservoir computer to capture the object's response to a fixed shaking input. This method allows the recovery of hidden properties, such as the center of mass, through a lightweight readout of the dynamics. The approach demonstrates a strategy to convert brief physical interactions into actionable cues for robotic perception.
ForeTac-VLA is a forecasting-based model designed for tactile-vision-language-action in contact-rich robotic manipulation. It addresses the limitation of existing models by explicitly predicting future tactile states to improve robustness in physical interactions. The model fuses tactile observations with vision-language features via cross-attention and uses a transformer to forecast multi-step tactile states. This fused information guides action generation within a VLA backbone, utilizing a ground-truth-to-prediction curriculum for stable training. The method achieved a 95% success rate across four tasks, significantly outperforming other VLA methods.
SCOUT is a ROS-integrated framework for real-time aerial defense against unauthorized unmanned aerial vehicles by onboard UAVs. It performs visual detection, target association, track filtering, and control command generation without external sensing. The system uses TensorRT for drone detection and ByteTrack for association, incorporating a state machine to ensure stable control by rejecting target jumps. Evaluation showed the system maintained valid target detections for 92.2\% of frames during a planar interception strategy. This demonstrates the feasibility of visual tracking and closed-loop control for UAV perimeter defense.
SPARROW is a belief-space planner based on Partially Observable Monte Carlo Planning (POMCP) designed for adaptive robot routing in the presence of temporary obstacles. It models navigation as a partially observable semi-Markov decision process, searching over traversal, observation, and waiting actions while maintaining particle beliefs about obstacle classes and clearance times. The planner learns survival models online from observations and simulates obstacle arrivals using a generative model. A value-of-learning criterion is introduced to balance data collection against reducing future navigation regret. SPARROW demonstrated performance improvements, reducing mean time-to-goal by 12-26% in simulations and 20.5% on physical robots compared to OSCAR.
Improvements to Visual-Inertial SLAM are crucial for making advanced marine robotics accessible to broader applications. This work demonstrates that robust navigation is possible on low-cost Autonomous Underwater Vehicles (AUVs) using only inexpensive, consumer-grade sensors. The research evaluates stereo cameras, MEMS IMUs, and depth sensors in a 6-DOF underwater environment. Optimizations are proposed for sensor fusion to handle the visual and physical challenges of unconstrained underwater operation. The findings establish an accessible benchmark for real-time maritime missions without requiring expensive industrial-grade hardware.
Vision-Language-Action (VLA) models are used in robotic manipulation but struggle with responsiveness when execution conditions change. The VLA-Feedback approach uses a two-timescale architecture combining low-frequency diffusion planning with high-frequency visual feedback. This method retains a lightweight feedback interface for real-time action correction using the latest observations before execution. This design avoids rerunning the full vision-language diffusion model for each correction. The technique improved success rates on dynamic simulation and real-robot tasks compared to baseline methods.