Researchers from the Oxford Robotics Institute and Georgia Tech have published ScaRF-SLAM, a framework that separates camera tracking from dense mapping by pairing classical feature-based SLAM with geometric foundation models. Benchmarks show reconstruction errors as low as 2 cm per 10 m indoors and 10 cm per 30 m outdoors against LiDAR ground truth.
Researchers have demonstrated the Adversarial Calibration Attack (ACA), the first known physical attack to subvert online camera-LiDAR calibration in autonomous vehicles using a single printed poster. By inducing up to 33.9° of rotational calibration error, ACA propagates system-wide failures from perception through to planning and control.
Researchers have built FleetScape, a mixed reality sandtable interface that lets a single operator supervise up to 15 drones through layered spatial data and fluid control-mode switching. A user study with six professional pilots shows the system boosts situational awareness — but also exposes a ceiling effect as fleet size grows.
Researchers have released PIVOT, a multi-trajectory dataset captured with a DJI Mini 4 Pro that systematically tests NeRF and 3D Gaussian Splatting methods under realistic drone-operating conditions. Benchmark results reveal consistent quality gaps when models are evaluated on camera paths and pose sources not seen during training—exposing a fundamental evaluation blind spot in the field.
A new arXiv survey unifies five previously isolated research communities—privacy filters, unlearnable examples, generative safeguards, adversarial CAPTCHAs, and provenance mechanisms—under a single 'adversarial attacks for good' paradigm. The authors find that most protections remain validated only against static or weakly adaptive adversaries, leaving a critical deployment gap as visual AI pipelines evolve toward multimodal agents.
A new arXiv preprint introduces a zero-shot framework that estimates conflict-zone building impacts using only archival maps and LLM-extracted weapon data — no post-strike satellite imagery required. The system pairs Hopkinson-Cranz blast-radius physics with adaptive 2D segmentation and depth-augmented Large Vision-Language Models to count exposed structures in both sparse and dense urban environments.
Researchers at Peking University have published LabEvolver, a training-free framework that gives robotic wet-lab agents persistent episodic memory by distilling execution trajectories into reusable skill and safety experience. In real-world solution-preparation trials, it cut pH-regulation task time by 48.2% and reduced safety-gate intercepts by 60.0%, while on the ALFWorld benchmark it lifted cumulative success from 76.2% to 91.4% across 500 continual tasks.
A new arXiv paper, DiffAttack, uses latent diffusion model optimization to generate adversarial faces that fool deep face recognition systems at an 84.86% average success rate across multiple models. The framework significantly outperforms both noise-based and semantic adversarial methods, raising urgent questions for operators deploying UAV-mounted biometric identification systems.
Researchers have identified a mechanism called policy-critical action-to-vision attention hijacking, through which printable adversarial patches can drive a VLA robot's failure rate to 100%. Their proposed defence, SARF, cuts that failure rate to as low as 14.2% with zero added inference overhead.
Researchers have published GPAC, a four-layer hierarchical control architecture that allows an arbitrary number of quadrotors to cooperatively transport a cable-suspended payload without any central coordinator, shared payload mass data, or inter-agent cable-state exchange. High-fidelity simulation across 13 randomised trials yielded a mean payload-tracking error of just 33.8 cm, with all control and estimation loops closed through onboard sensors alone.
Researchers have trained a recurrent neural network whose architecture is directly wired from the synaptic-resolution connectome of the fruit fly Drosophila melanogaster to perform vision-based robot navigation. FLYNN outperforms conventional hand-crafted networks on out-of-distribution scenarios and keeps functioning even under total camera blackout — without any retraining.
Researchers have trained a deep reinforcement learning policy that enables a quadrotor to accurately throw a cable-suspended payload to a designated target, cutting landing error by up to 50% and throw duration by up to 30% versus model-based baselines. The policy transfers zero-shot from simulation to real hardware, and a companion vision-driven variant matches the accuracy of the state-based version.