Autonomous Systems: From Robot Safety to Smart Medical Predictions, and Everything In-Between
Latest 9 papers on autonomous systems: Sep. 27, 2026
Autonomous systems are rapidly evolving, promising transformative advancements across industries from healthcare to robotics. However, developing these systems presents complex challenges, including ensuring safety, enhancing human-robot collaboration, optimizing resource utilization, and enabling robust performance in unpredictable environments. Recent research in AI/ML is tackling these hurdles head-on, delivering innovative solutions that push the boundaries of what autonomous agents can achieve.
The Big Idea(s) & Core Innovations
One central theme in recent advancements is improving the dependability and safety of autonomous agents. A groundbreaking study by Jasper J. van Beers and his colleagues from the Faculty of Aerospace Engineering, Delft University of Technology introduces a novel approach to predicting instability in autonomous control systems. Their paper, “Indicators of resilience for autonomous control systems,” demonstrates how critical slowing down (CSD) can serve as a model-free early warning signal for impending instability, even in complex systems like quadrotors with damaged propellers. This is crucial because it allows for proactive intervention before catastrophic failure, a significant leap for robot safety. Complementing this, in the realm of AI safety, research from Alizishaan Khatri, Chiquita Prabhu, and Omkar Neogi at Wrynx Inc. and independent researchers, in their paper “Safety Beyond the Interface: Detecting Harm via Latent States in Large Language Models,” shows that safety-relevant information is already encoded within the hidden states of Large Language Models (LLMs). They’ve developed lightweight MLP probes that can detect harmful content with remarkable accuracy and zero latency, offering a fundamental shift in how we build safer LLM-powered applications.
However, this newfound power of LLMs also introduces new vulnerabilities. Doniyorkhon Obidov and his team from the Electrical and Computer Engineering Department, Michigan Technological University unveil a chilling new threat in their paper, “Silent Sabotage: Internal State Triggered Backdoor Attacks on LLM-Powered Robotic Systems.” They demonstrate that LLM-powered robots can be backdoored by manipulating instructions such that the robot’s own sequence of past actions acts as a trigger, achieving near-perfect attack success rates while remaining stealthy. This highlights a critical gap in existing defenses focused on external stimuli and emphasizes the need for internal state-aware security measures.
Beyond safety, optimizing human-robot interaction and resource management is a key focus. Rohith Prem Maben and Ayesha Jena from Lund University present an interactive simulation platform for “Towards Intent-Aware Human-Robot Teaming: A Platform for Search-and-Rescue Operations.” By applying the Joint Control Framework (JCF), they model operator intent across multiple cognitive levels, enabling adaptive decision support for UAV-UGV coordination in search-and-rescue (SAR) missions. This work helps understand how humans dynamically allocate functions with autonomous systems, paving the way for more intuitive and effective collaboration.
For resource optimization, especially in edge computing, Mahdi Taheri and colleagues from Humboldt University of Berlin and Tallinn University of Technology introduce WARD, a “Runtime Workload-Adaptive Vision TRansformer Framework for Dependable Edge AI.” WARD ensures reliable inference under dynamic power budgets and reliability requirements by combining channel-wise subnetwork partitioning, reliability-aware continual learning, and dynamic operating-mode scheduling. This allows autonomous systems to adapt their computational footprint without sacrificing dependability. Similarly, in underwater remote sensing, Logan Luna and his team from Georgia Institute of Technology and Embry-Riddle Aeronautical University propose a “Signal-Centric Remote Sensing via Alternative Preprocessing and Acoustic Processing for ML-Driven Applications.” Their novel CSV-based sonar data processing bypasses computationally intensive image processing, achieving a dramatic 91.18% reduction in processing time while improving ML-based object detection accuracy for autonomous underwater vehicles.
Finally, the very architecture of autonomous agents is being re-evaluated. Yishuo Yuan and Jiaheng Liu from Nanjing University and Shanghai Jiao Tong University challenge the common assumption that more agents are always better in their paper, “Rethinking Multi-Agent Collaboration: When More Is Less.” Through graph-theoretic analysis, they show that multi-agent collaboration is truly beneficial only for long-horizon tasks with sparse dependencies, while single-agent systems excel in tightly coupled sequential workflows. This insight is crucial for designing efficient multi-agent systems. Further enhancing control strategies, Giray Önür and his team from the Delft Center for Systems and Control, Delft University of Technology introduce “Composite-Gradient Learning for Shared Control Authority Between Deep Reinforcement Learning and Model Predictive Control.” This novel method integrates Model Predictive Control (MPC) into Deep Reinforcement Learning (DRL) by representing their control inputs as a joint action, allowing the DRL agent to account for both direct and indirect effects, leading to more robust control policies, especially in complex systems like freeway traffic networks.
From a medical perspective, Erik Aerts and his collaborators from Chalmers University of Technology and University of Gothenburg present TRACER in their paper “AI-based prediction of worsening heart failure from low-resolution telemonitoring data.” TRACER uses a Transformer with Contrastive Event Representation to predict heart failure hospitalization from sparse, low-resolution telemonitoring data, achieving impressive sensitivity by reformulating prediction as event detection within time windows and utilizing anomaly-based contrastive pre-training.
Under the Hood: Models, Datasets, & Benchmarks
These advancements are powered by innovative models, specialized datasets, and rigorous benchmarks:
- TRACER Model: A Transformer with Contrastive Event Representation, specifically designed by Chalmers University of Technology for rare event prediction in medical time series, leveraging time-aware embeddings and anomaly-based contrastive pre-training. Code available at https://github.com/aeerik/TRACER.git.
- LLaMA-3.1-8B: Heavily utilized by Wrynx Inc. for safety probing, demonstrating that its hidden states contain linearly separable safety signals. Tested against WildJailbreak, Beavertails, and AEGIS 2.0 datasets.
- SAIGE (Semantic-Aware Incremental Graph Evolution): A lightweight multi-agent collaboration mechanism proposed by Nanjing University that models collaboration as a dynamically evolving graph. Evaluated on Terminal Bench 2.1, NL2Repo Bench, Deep Research Bench II, and AgentIF-OneDay benchmarks using DeepSeek-V4-Pro.
- WARD Framework: A runtime-adaptive Vision Transformer framework from Humboldt University of Berlin that uses channel-wise subnetwork partitioning and reliability-aware continual learning. Validated on EuroSAT and ImageNet-21K pretrained weights, with a lightweight FPGA implementation.
- ArduPilot SITL and Unity: The core of the SAR simulation platform by Lund University for UAV-UGV coordination, utilizing the MAVLink protocol for real-time telemetry. Code available at https://github.com/jayesha94/IA-HRI.git.
- CSV-based Sonar Processing: A novel data processing method for Ping 360 sonar data (BlueRobotics) proposed by Georgia Institute of Technology that avoids image conversion for autonomous underwater vehicles, improving ML predictions.
- Composite-Gradient Learning (CGL): A hierarchical DRL-MPC control framework developed by Delft University of Technology for shared control authority, specifically evaluated on multi-class freeway traffic networks. Code available at https://github.com/GirayOnur/composite-gradient-learning.
- CSD-based Resilience Indicators: Utilizes actuator response measurements to monitor stability degradation in autonomous control systems. Validated on simulations of an inverted pendulum and quadrotor, and real-world quadrotor flight data with damaged propellers. Code available at https://github.com/tudelft/indiflight.
Impact & The Road Ahead
These advancements herald a new era for autonomous systems. The ability to predict system instability without models, detect internal LLM harm with zero latency, and develop intent-aware human-robot teams will dramatically enhance the safety, reliability, and effectiveness of robots and AI agents in critical applications. The insights into efficient multi-agent design and novel data processing for edge and underwater environments promise to unlock new capabilities and optimize resource usage.
The findings on internal state-triggered backdoor attacks on LLM-powered robots are a stark reminder of the urgent need for robust cybersecurity frameworks tailored to the unique vulnerabilities of highly autonomous systems. Future research will undoubtedly focus on developing countermeasures that address these ‘silent sabotages.’ Furthermore, the integration of DRL and MPC through composite gradients points towards increasingly sophisticated and robust control policies for complex real-world challenges. As we move forward, the convergence of robust safety mechanisms, efficient collaboration paradigms, and adaptive intelligence will undoubtedly lead to more intelligent, dependable, and impactful autonomous systems that seamlessly integrate into our world.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment