Autonomous Systems: From Risk-Aware AI to Sensing Shadows and Shared Control
Latest 11 papers on autonomous systems: Jun. 27, 2026
Autonomous systems are rapidly evolving, promising transformative changes across industries from robotics and self-driving cars to scientific discovery and assistive technologies. But this progress brings complex challenges: how do we ensure these systems are reliable, safe, and truly intelligent? Recent research dives deep into these questions, offering groundbreaking solutions in areas like robust perception, human-AI collaboration, theoretical foundations, and even the very definition of intelligence.
The Big Idea(s) & Core Innovations
At the forefront of robust perception, the paper Beyond a Shadow of a Doubt: Close Proximity Geometry Reconstruction Using FMCW Radar Shadow Effects by Felix de Trogoff du Boisguezennec, Benjamin Ramtoula, and Daniele De Martini from Oxford Robotics Institute, unveils a surprisingly simple yet powerful innovation. They demonstrate how vehicle chassis naturally creates consistent radar shadows. These shadows, typically dismissed as noise, can be exploited as a stable geometric reference cue. By analyzing the ratio of inner to outer radar return boundaries, they achieve a closed-form analytical mapping to estimate the inclination of nearby vertical objects with impressive accuracy (sub-degree in simulation, 3-4 degrees in real experiments). This is a game-changer for extending 2D radar perception into 3D reconstruction, especially for unstructured terrains where other sensors might struggle.
Complementing this perceptual advancement, ESMStereo: Enhanced ShuffleMixer Disparity Upsampling for Real-Time and Accurate Stereo Matching from Mahmoud Tahmasebi and colleagues at Atlantic Technological University, tackles the critical need for real-time, accurate depth estimation. Their Enhanced ShuffleMixer (ESM) module efficiently upsamples low-resolution disparity maps while recovering fine scene details, achieving remarkable speeds (116 FPS on RTX 4070S, 91 FPS on Jetson AGX Orin) on KITTI benchmarks. This innovation makes high-quality stereo matching viable for edge devices in autonomous systems by shifting computational load from expensive 3D operations to efficient 2D modules.
In the realm of human-AI collaboration, the “one body, two minds” paradigm emerges with One Body, Two Minds: Variable Autonomy Approach for a Co-embodied Robotic Hand by Piotr Koczy et al. from KTH Royal Institute of Technology. This paper introduces co-embodiment where a human and a wearable robotic hand share a single physical body and operate at variable autonomy levels. Users provide intent and position the hand, while a learning-from-demonstration visuomotor diffusion policy autonomously grasps objects. A user study showed a 23.3% improvement in task completion times and high acceptance, highlighting a powerful new model for assistive robotics where human and robot strengths are synergistically combined.
Beyond just physical interaction, AI is also transforming cognitive science. Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist by Akshay K. Jagadish and his team at Princeton University, introduces AUTOCOG. This fully autonomous agentic AI system uses LLM agents to propose competing theories, design discriminating experiments, collect behavioral data from online participants, and iteratively refine theories. AUTOCOG not only recovered ground-truth decision-making strategies but also discovered a novel “Diminishing Returns WADD” theory, demonstrating that data-driven discovery can overcome inherent LLM biases.
Finally, laying down theoretical groundwork for more robust AI, Risk-Aware Information Theory by Hamidou Tembine from UQTR, fundamentally redefines information theory. By replacing the classical expectation operator with expectiles, it introduces risk-sensitive measures like expectile entropy and mutual information. This new framework reveals properties impossible under Shannon’s risk-neutral approach, such as negative divergence under risk-seeking regimes, and provides a crucial lens for understanding and building truly “Superintelligent” systems that can account for tail risks, a limitation Shannon Intelligence inherently possesses.
Under the Hood: Models, Datasets, & Benchmarks
The papers leverage and contribute to significant resources, pushing the boundaries of autonomous systems:
- Radar Shadow Geometry: validated with RadaRays radar simulator and real-world experiments using a Navtech CTS350-X radar and the Oxford Offroad RobotCar Dataset (OORD) for real-world validation. This novel approach introduces chassis shadows as a new geometric cue for 3D reconstruction.
- ESMStereo: Evaluated extensively on KITTI 2012 and KITTI 2015 benchmarks, SceneFlow (synthetic), ETH3D, and Middlebury 2014 datasets. The key innovation, the Enhanced ShuffleMixer (ESM) module, is also shown to be portable and enhance existing architectures like PSMNet and Fast-ACVNet-Plus. Code is available at https://github.com/M2219/ESMStereo.
- Co-embodied Robotic Hand: Utilizes a learning-from-demonstration visuomotor diffusion policy for autonomous grasping and hands-free head gesture control. The project’s resources are available at https://co-embodiment.github.io/.
- AUTOCOG: While the system itself is the core innovation, the iterative discovery loop involves collecting behavioral data from online participants and validating findings in preregistered studies (e.g., https://osf.io/f7kes).
- Agentic Web Browsers: Evaluated using Perplexity Comet’s Voice User Interface, highlighting the practical application of Large Language Models (LLMs) for assistive technologies.
- Reliable Autonomous Systems: The ERAS workshop report outlines a comprehensive research roadmap, referencing open-source tools like ROSMonitoring (runtime verification), FRET (requirements formalization), Dafny (static verification), AADL (architectural modeling), RoboChart (software modeling), RoboSim (physical modeling), Gazebo (simulation), and ROS 2 (robotics framework). These tools are critical for the verification and validation (V&V) of complex autonomous systems. More details at https://erasrobotics.org.
- Risk-Aware Information Theory: A theoretical framework built upon the novel use of expectiles instead of classical expectations, offering a new mathematical foundation.
- BLENDS: The Bayesian Learning-Enhanced Deep Smoothing framework for GNSS-denied navigation leverages a transformer-based learning module within a classical Bayesian smoothing framework. It’s validated on the INSANE benchmark dataset for UAV navigation.
- Logical Compliance: The Rule Violation Score (RVS) is introduced as a metric for evaluating logical compliance on datasets like Family, FB15k-237 (knowledge graphs), and DV3F (relational database). Code can be found at https://anonymous.4open.science/r/Rule-Violation-Score-585C.
- LiDAR Security: Research demonstrates attacks on real tactical unmanned ground vehicles using sensor-internal malware, emphasizing vulnerabilities in LiDAR firmware.
- Accountability in Drone Firefighting: Investigated through real-life field trials with human participants, assessing drone roles within hierarchical structures.
Impact & The Road Ahead
These advancements collectively pave the way for more capable, reliable, and human-centric autonomous systems. The ability to extract 3D information from 2D radar (Beyond a Shadow of a Doubt: Close Proximity Geometry Reconstruction Using FMCW Radar Shadow Effects) and achieve real-time accurate stereo depth (ESMStereo: Enhanced ShuffleMixer Disparity Upsampling for Real-Time and Accurate Stereo Matching) means smarter and safer navigation for autonomous vehicles and robotics, especially in challenging environments. The concept of co-embodiment (One Body, Two Minds: Variable Autonomy Approach for a Co-embodied Robotic Hand) opens new avenues for assistive robotics, enhancing human capabilities through intuitive, shared control.
The groundbreaking work with AUTOCOG (Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist) highlights AI’s potential as a creative scientific partner, accelerating discovery in fields like cognitive science and beyond. However, as autonomous systems become more intelligent and integrated, the need for robust verification and accountability becomes paramount. The Engineering Reliable Autonomous Systems: Challenges and Solutions workshop report emphasizes that most deployed systems operate under constraints, highlighting the critical role of Operational Design Domains (ODDs) and the need for combining diverse verification techniques to build confidence. Furthermore, the Accountability in Autonomous Drone-Based Firefighting: Insights From a Field Trial study reminds us that technological advancements must be matched with clear organizational structures and human-centered design to ensure responsible deployment.
Security is another non-negotiable aspect, as demonstrated by Anywhere, Any-Stymie: Remote Activation of Trojan Malware on LiDAR with Modulated Signals, which uncovers a severe new attack surface in LiDAR firmware, urging the need for sensor-level integrity guarantees. Finally, the introduction of a Risk-Aware Information Theory provides a profound theoretical underpinning for understanding and building truly robust and intelligent systems that can navigate uncertainty and extreme risks, while the Rule Violation Score (RVS) offers a crucial metric to ensure that even highly accurate predictive models adhere to logical constraints, essential for high-stakes applications like healthcare and finance.
The trajectory is clear: autonomous systems are moving towards increased sophistication, adaptability, and integration with human users. The next steps involve bridging the gap between symbolic and sub-symbolic AI through neuro-symbolic approaches, as highlighted by the ERAS report, enhancing explainability and trustworthiness, and developing comprehensive standards for certification. The future of autonomous systems promises not just efficiency, but a deeper, more reliable, and safer collaboration between humans and machines.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment