Data from Ukrainian drone flights is being converted into a valuable resource for the defense sector and commercial companies. This battlefield data allows AI developers to train models using real-world, chaotic conditions that are difficult to replicate in labs. This process creates a market where combat experience becomes a commercial asset, fueling explosive growth. There is a need for regulation to govern how this data flows from combat to civilian applications and to address issues of consent and potential exploitation. Governments must establish frameworks to oversee the use of battlefield data to ensure accountability and protect individuals.
NASA's Cooperative Autonomous Distributed Robotic Exploration (CADRE) mission plans to send three rovers to the Moon to autonomously explore a patch of ground. These rovers will operate as a self-guided team, electing leaders and redistributing tasks without direct human control. This testing is crucial for future lunar missions, such as accessing water ice for a lunar economy. The mission addresses significant challenges on the Moon, including communication blackouts and extreme thermal conditions. Success in this test will pave the way for more complex, long-duration robotic operations on the lunar surface.
A robotic ball called RoboBall is being developed to explore inaccessible lunar terrain like the Shackleton Crater. The design utilizes an inflatable sphere that rolls by shifting its center of mass via a pendulum mechanism, offering unique advantages for low-gravity environments. This concept allows the ball to potentially collect samples and launch them back to a rover using small rockets. Although the current prototype has limitations regarding material science and mission readiness, the team plans further engineering. The research aims to advance the collection of lunar geological data for future missions.
Domestic robot perception exports can still leak sensitive household information through semantic and spatial representations, even when raw sensor data is kept local. The authors introduce Task-Functional Perception Distillation (TFPD) to create a framework that exports representations based on task utility and inferred privacy risks. Evaluations show that while achieving identical task success metrics, the exported representations still have measurable linkability to private information. Further analysis indicates that removing fields or increasing abstraction does not establish a universal privacy ordering for these exports. Consequently, the research advocates for task-specific, multi-risk evaluation of the complete public robot representation.
Training Vision-Language-Action (VLA) policies on limited demonstrations can lead to modality entanglement, causing policies to over-rely on spurious sensor correlations. Evidence-Gated Regularization (EGR) is proposed as a modality-agnostic objective to address this issue without inference-time overhead. EGR introduces task-relevance signals to gate consistency objectives, ensuring invariance to low-evidence sensors and sufficiency based on high-evidence ones. Validation on real-robot setups showed significant improvements in success rates when facing sensor corruption or distractors. Overall, EGR substantially enhances robustness across different embodiment types.
A Koopman-based model predictive control framework using global and local observables is presented for real-time shape control of multi-segment soft robotic arms. This approach addresses the challenge of controlling continuum deformation by incorporating segment coupling, loading, and inertial effects. Numerical experiments confirm the controller's scalability up to ten actuated segments. Physical experiments demonstrated real-time shape control for 3- and 5-segment arms with robust tracking and disturbance recovery. The results show the framework enables dynamic, accurate, and scalable shape control.
The paper introduces the Central-Path Certifier (CP-Cert), a certifiable method designed to certify optima for problems with degeneracy in convex semidefinite programming relaxations. CP-Cert works by finding a nearby feasible region, or central path, to obtain a valid certificate efficiently. The method leverages indirect linear algebra techniques, problem sparsity, and parallelism to maintain efficiency. It is applied to matrix-weighted pose registration and pointcloud data association, showing runtimes up to three orders of magnitude faster than direct solvers in simulations. Ultimately, the work results in a certifiable, outlier-robust pose-estimation pipeline for real-world data.
StyleDrive is a world-model-based learning framework designed for multi-style end-to-end autonomous driving. It addresses limitations in existing methods by enforcing long-horizon consistency through temporal regularization and disentangling ego-environment interactions. The framework employs Group Relative Policy Optimization to enable multi-style driving behaviors without retraining. Evaluation on Bench2Drive achieved a driving score of 88.44 and a success rate of 66.82. StyleDrive also demonstrated promising sim-to-real transfer capabilities on real vehicle platforms.
A multimodal dataset is proposed to analyze user engagement in human-robot interaction (HRI). This dataset is constructed by integrating wearable physiological signals, behavioral data, and self-report measures. The data collection protocol is structured across different levels of task complexity in an experiment. This approach addresses the limitation of prior work focusing primarily on observable cues. The goal is to create a comprehensive dataset for understanding human engagement with humanoid robots.
R2S-Eval is an evaluation pipeline that combines real-to-sim calibration with vision-language model preference evaluation for robot manipulation policies. It uses real-to-sim calibration to generate simulator rollouts, reducing the need for repeated hardware trials. A Vision-Language Model (VLM) evaluator assesses execution quality by generating pairwise preferences from these videos. These preferences are aggregated into policy rankings, which are further validated against human preferences. Experiments show that R2S-Eval yields stable conclusions, aligns with human judgments, and reveals quality-based behavior differences.
ARTiS is a novel gripper designed for enhanced tool manipulation during disassembly applications. It integrates the adaptability of soft grippers, the dexterity of anthropomorphic hands, and the robustness of rigid mechanisms. The design features a soft palm and fingertips utilizing active jamming and fin-ray adaptation for secure tool holding. High finger dexterity is achieved through a seven degrees of freedom design, allowing fingertips to orient to any surface. Evaluations confirmed the gripper's compliance, durability, and functional versatility when handling various disassembly tools.
The research addresses the need for agility and robustness in robotic manufacturing and logistics by integrating robot skills. It proposes an easy-to-use system for programming and executing parametrized skills and skill monitors based on CAD models. The execution control is dynamic and parametrized, utilizing a modified Behavior Tree with event-based communication. A human-robot-crane collaboration example demonstrates a human instructing an overhead crane and manipulator to insert an object. This system facilitates collaborative tasks by providing a mechanism for robots to perform varied tasks with specified parameters.