Wednesday, October 14, 2026

Could AI become the missing mediator between drivers and automated vehicles?

Tsinghua University Press

Evolution of Human–Vehicle Collaboration in Automated Driving — image:
Illustration of the proposed Mediator-in-the-Loop-Driving (MILD) paradigm. The framework introduces an AI-based mediator between the driver and the automated driving system, enabling bidirectional understanding and collaborative decision-making through joint perception of driver states and traffic environments.
view more
Credit: Communications in Transportation Research

As vehicles become increasingly automated, the role of human drivers is changing. Instead of continuously controlling the vehicle, drivers are often expected to supervise automated systems and intervene when necessary. However, this transition has created a new challenge: drivers frequently struggle to understand why a vehicle behaves in a certain way, while automated systems have limited awareness of the driver’s changing cognitive state, emotions, and preferences.

To address this challenge, researchers from McGill University, The Hong Kong University of Science and Technology (Guangzhou), and the University of Luxembourg have developed a new human-vehicle collaboration framework called Mediator-in-the-Loop-Driving (MILD). Their study proposes an intelligent agent that acts as a mediator between the driver and the automated driving system, helping both sides maintain a shared understanding of the driving situation.

The team published their study in Communications in Transportation Research (https://doi.org/10.26599/COMMTR.2026.9640038).

“Current automated vehicles largely rely on a supervision paradigm, where humans monitor automation but automation has only limited understanding of the human,” said Jiangbo Yu, corresponding author of the study. “We envision a future where intelligent agents help drivers and automated systems collaborate as teammates rather than operate as separate entities.”

From supervision to collaboration

Most existing automated driving systems focus primarily on understanding the surrounding traffic environment. While driver monitoring systems can detect fatigue or distraction, they are often isolated warning systems that do not actively contribute to collaborative decision-making.

The MILD framework introduces a new paradigm in which an AI-powered mediator continuously interprets both the driver’s state and the external driving environment. Instead of treating drivers as passive supervisors, the system helps them become active managers of the driving process by providing personalized recommendations, explanations, and assistance strategies.

The researchers argue that this bidirectional understanding is essential for future intelligent vehicles, especially in partially automated driving scenarios where both humans and automation share responsibility for safety.

“Our goal is not to replace human drivers,” explained Jiyao Wang, the first author of the study. “Instead, we want to create a system that helps drivers and automated vehicles understand each other and make better decisions together.”

Combining driver understanding with traffic awareness

A key innovation of MILD is its ability to jointly analyze information from both inside and outside the vehicle.

The system integrates driver-related information, such as attention, emotional state, and driving behavior, with external traffic information including surrounding vehicles, road conditions, and potential hazards. This joint perception enables the AI agent to develop a more comprehensive understanding of the driving situation than conventional vehicle-centric systems.

Based on this understanding, a lightweight large language model generates high-level recommendations, including driving suggestions, safety warnings, human-machine interaction strategies, and personalized cockpit adjustments. Importantly, the system does not directly control steering, braking, or acceleration, ensuring that the driver remains an active participant in the decision-making process.

Ensuring trustworthy and explainable AI decisions

One of the major concerns surrounding AI-assisted driving is whether the generated recommendations are safe, reliable, and aligned with human expectations.

To address this issue, the research team developed a new alignment framework called Evidence- and Constraint-weighted Policy Optimization (ECPO). The method evaluates AI-generated recommendations against multiple layers of constraints, including traffic regulations, vehicle capabilities, driver preferences, and contextual evidence from the driving environment.

This mechanism helps ensure that recommendations are not only accurate but also explainable, auditable, and consistent with real-world driving requirements.

“Future intelligent vehicles will need more than automation,” said Jiyao Wang. “They will require transparent communication and mutual understanding between humans and machines. We see MILD as an important step toward that vision.”

Toward next-generation intelligent cockpits

The researchers evaluated MILD using multiple public datasets covering driver monitoring, traffic-scene understanding, and advanced driver assistance applications. Experimental results demonstrated improvements in both perception performance and strategy quality, while human evaluators rated the generated recommendations as safer, more comfortable, and more explainable than existing baseline approaches.

The team believes that intelligent agents will become a key component of future intelligent cockpits, helping bridge the communication gap between humans and automated systems. Beyond automated driving, the proposed framework could also contribute to broader human-AI collaboration scenarios where trust, transparency, and shared decision-making are essential.

About Communications in Transportation Research

Communications in Transportation Research was launched in 2021, with academic support provided by Tsinghua University and China Intelligent Transportation Systems Association. The Editors-in-Chief are Professor Xiaobo Qu, a member of the Academia Europaea from Tsinghua University, and Professor Xiaopeng (Shaw) Li from University of Wisconsin–Madison. The journal mainly publishes high-quality, original research and review articles that are of significant importance to emerging transportation systems, aiming to serve as an international platform for showcasing and exchanging innovative achievements in transportation and related fields, fostering academic exchange and development between China and the global community.

It has been indexed in SCIE, SSCI, Ei Compendex, Scopus, CSTPCD, CSCD, OAJ, DOAJ, TRID and other databases. It was selected as Q1 Top Journal in the Engineering and Technology category of the Chinese Academy of Sciences (CAS) Journal Ranking List. In 2022, it was selected as a High-Starting-Point new journal project of the “China Science and Technology Journal Excellence Action Plan”. In 2024, it was selected as the Support the Development Project of “High-Level International Scientific and Technological Journals”. The same year, it was also chosen as an English Journal Tier Project of the “China Science and Technology Journal Excellence Action Plan PhaseⅡ”. In 2024, it received the first impact factor (2023 IF) of 12.5, ranking Top1 (1/58, Q1) among all journals in "TRANSPORTATION" category. In 2026, its 2025 IF was announced as 12.7, maintaining the Top1 position (1/66, Q1) in the same category.

From Volume 6 (2026), Communications in Transportation Research will be published by Tsinghua University Press on the SciOpen platform with the official journal website at https://www.sciopen.com/journal/2097-5023. We kindly request that all new manuscript submissions be made through the journal’s submission system at https://mc03.manuscriptcentral.com/commtr. For any submission-related inquiries, please contact the Editorial Office at commtr_e@mail.tsinghua.edu.cn.

Journal

Communications in Transportation Research

DOI

10.26599/COMMTR.2026.9640038

Article Title

MILD: Mediator Agent System with Bidirectional Perception and Multi-Layered Alignment for Human-Vehicle Collaboration

Autonomous vehicles learn to cooperate in complex traffic

Tsinghua University Press

COIN Framework Overview — image:
The overall COIN framework, which follows the CTDE paradigm. During training, an individual critic learns individual Q-values to guide the ego vehicle’s navigation policy, while a centralized critic learns global Q-values to facilitate collaborative policy learning. These two critics work together to achieve joint policy optimization. During execution, the actor learns an end-to-end policy based only on local observations. The right side illustrates the roles of the individual critic and centralized critic. Notably, we design a dual-level interaction-aware centralized critic with a counterfactual baseline to improve credit assignment and promote better agent cooperation.
view more
Credit: MARMot Lab, National University of Singapore

Autonomous vehicles in future intelligent transportation systems need to do more than navigate safely on their own. In complex scenarios such as unsignalized intersections, roundabouts, and bottleneck roads, they must also coordinate with surrounding vehicles to improve overall safety and efficiency. However, existing end-to-end driving methods mainly focus on individual navigation, while many multi-agent reinforcement learning methods only model local or neighborhood interactions. This makes it difficult to jointly optimize individual driving performance and system-level cooperation.

To address this challenge, researchers from the National University of Singapore, The Hong Kong Polytechnic University, and Tsinghua University developed COIN, a collaborative interaction-aware reinforcement learning framework for multi-agent self-driving systems. COIN follows a centralized training and decentralized execution paradigm. During training, vehicles learn cooperative strategies using global information; during execution, each vehicle makes decisions based only on its local observations, making the framework suitable for decentralized autonomous driving.

The core of COIN is counterfactual individual-global TD3, or CIG-TD3 algorithm, which jointly optimizes individual navigation and global collaboration objectives. A counterfactual baseline is introduced to estimate each vehicle’s contribution to the overall system performance, helping the learning algorithm assign credit more effectively. COIN also uses a dual-level interaction-aware centralized critic. Conditional variational autoencoders capture pairwise interactions between vehicles, while graph attention networks model system-level dependencies across the traffic scene.

The framework was evaluated in MetaDrive under three dense traffic scenarios: unsignalized intersections, roundabouts, and bottleneck roads. COIN outperformed IPPO, ITD3, CoPO, and TraCo in success rate, collision rate, safety, and efficiency. In the intersection scenario, COIN achieved an 88.78% success rate and reduced the collision rate to 3.94%. In the bottleneck scenario, it achieved a 96.33% success rate with only a 1.57% collision rate.

COIN was further validated on a hybrid real-simulation robotic platform. In a mock unsignalized intersection, eight Mecanum-wheeled mobile robots followed trajectories generated by COIN, while other vehicles were represented as virtual agents in simulation. This setup demonstrated that COIN can support safe and efficient cooperative navigation in complex multi-robot interaction scenarios.

By shifting multi-agent autonomous driving from independent decision-making toward cooperative optimization, COIN provides a promising direction toward safer and more efficient autonomous driving in dense urban traffic.

The team published their study in Communications in Transportation Research (https://doi.org/10.26599/COMMTR.2026.9640031).

About Communications in Transportation Research

Journal

Communications in Transportation Research

DOI

10.26599/COMMTR.2026.9640031

Article Title

COIN: Collaborative Interaction-Aware Multi-Agent Reinforcement Learning for Self-Driving System

Turning Foundation models into “driving mentors”: found-RL enables efficient reinforcement learning for autonomous driving

Tsinghua University Press

image:
Overall framework
view more
Credit: Communications in Transportation Research

Researchers developed Found-RL, an open-source platform that integrates vision-language models into reinforcement learning for autonomous driving. By using an asynchronous batch inference framework, Found-RL allows large foundation models to provide semantic feedback without slowing down high-frequency simulation training. Experiments in CARLA show that lightweight reinforcement learning agents can achieve competitive driving performance while maintaining real-time policy inference.

The team published their study in Communications in Transportation Research (https://doi.org/10.26599/COMMTR.2026.9640027).

Reinforcement learning has become a promising approach for end-to-end autonomous driving because it enables agents to learn driving policies through interaction with simulated environments. However, conventional reinforcement learning often requires extensive trial and error and may lack semantic understanding in complex traffic scenarios. Vision-language models, on the other hand, can interpret driving scenes and provide high-level reasoning, but their heavy inference cost makes them difficult to use directly in fast, multi-environment reinforcement learning loops.

“Large vision-language models can provide useful driving knowledge, but directly calling them at every simulation step would make reinforcement learning training extremely slow,” says Sikai Chen, corresponding author of the study and assistant professor at the University of Wisconsin–Madison. “Found-RL is designed to bridge this gap by allowing foundation models to guide reinforcement learning in the background, without blocking the simulation process.”

Turning foundation models into scalable driving mentors

The core of Found-RL is an asynchronous batch inference framework. Instead of waiting for a vision-language model to respond at every environment step, the platform separates simulation rollout from model inference. Multiple simulation workers send driving observations and context information to a shared request queue, while an inference server processes these requests in batches and returns feedback asynchronously.

This design allows the reinforcement learning agent to continue interacting with the driving environment while the vision-language model works as a background feedback provider. The returned feedback is then matched to the correct training sample and stored in the replay buffer for learning.

“Found-RL treats the vision-language model as a scalable mentor rather than a real-time controller,” Chen explains. “The lightweight reinforcement learning policy still performs high-frequency control, while the foundation model provides semantic supervision during training.”

Learning from both action guidance and semantic rewards

On top of the platform, the research team implemented two types of vision-language-model-based supervision. The first is action guidance, where the model provides expert-like driving actions that help the agent explore more efficiently. The team designed two mechanisms, Value-Margin Regularization and Advantage-Weighted Action Guidance, to incorporate these action suggestions into reinforcement learning.

The second is CLIP-based reward shaping. To address the difficulty of applying CLIP directly to continuous driving control, the researchers introduced a conditional contrastive action-alignment reward. The method conditions text prompts on driving context, such as vehicle speed and navigation command, and compares the agent’s action with context-specific semantic action anchors. This provides dense and interpretable reward signals during training.

Lightweight agents with real-time inference

The team evaluated Found-RL in CARLA using several autonomous driving benchmarks, including Leaderboard and NoCrash. The results show that Found-RL improves the performance of classical reinforcement learning algorithms in terms of driving score, route completion, and safety-related metrics.

In the Leaderboard benchmark, the DrQv2-CLIP agent achieved a Driving Score of 0.77 and a Success Rate of 57%, outperforming standard reinforcement learning baselines. The study also found that vision-language-model-enhanced agents could reduce unsafe behaviors such as vehicle collisions and red-light violations in several settings.

A key finding is that Found-RL can transfer useful semantic knowledge from large foundation models into compact reinforcement learning policies. The resulting agent contains only 3.82 million parameters but can achieve driving performance competitive with much larger vision-language models. In addition, the lightweight policy reaches about 500 FPS in forward-pass inference, compared with approximately 1 FPS for several billion-parameter vision-language models.

“These results suggest that foundation models do not necessarily need to drive the vehicle directly,” says Chen. “They can instead serve as teachers during training, helping smaller and faster policies learn safer and more efficient driving behaviors.”

Toward foundation-model-enhanced autonomous driving

The study provides an open-source framework for future research at the intersection of foundation models, reinforcement learning, and autonomous driving. The authors note that future work may include testing Found-RL on more challenging driving benchmarks, accelerating and quantizing vision-language-model inference, using multiple foundation models to provide specialized feedback, and exploring lightweight vision-language models as policy backbones.

“Autonomous driving requires both fast control and rich scene understanding,” Chen says. “Found-RL offers a practical way to combine these strengths by using foundation models to improve training while keeping the deployed driving policy efficient.”

About Communications in Transportation Research

Journal

Communications in Transportation Research

DOI

10.26599/COMMTR.2026.9640027

Article Title

Found-RL: foundation model-enhanced reinforcement learning via asynchronous VLM feedback for autonomous driving

Can vision-language models understand unstructured road scenes under degraded sensing conditions?

Tsinghua University Press

Overview of the USB test and evaluation framework — image:
Overview of the test and evaluation framework. Scenes from unstructured roads, captured via multiple camera views, are expanded into twenty corruption types. These inputs are used to generate questions covering perception, prediction, explanation, and planning, which are then evaluated using a combination of objective and subjective metrics.
view more
Credit: Communications in Transportation Research

Can vision-language models understand unstructured road scenes under degraded sensing conditions? Researchers developed the Unstructured Scene Benchmark (USB), a test and evaluation framework for autonomous driving models in unstructured road scenes and degraded visual conditions. By testing seven open-source vision-language models across six camera views, 20 corruption types, and 45,969 question-answer samples, the study reveals uneven robustness, weak planning ability, and signs of insufficient visual grounding.

The team published their study in Communications in Transportation Research (https://doi.org/10.26599/COMMTR.2026.9640034).

Vision-language models (VLMs) are increasingly used to connect visual perception, language reasoning, and decision support in autonomous driving. However, many existing benchmarks focus on structured urban roads and relatively clean imagery. Real roads are often less orderly: lane boundaries may be unclear, temporary construction may change traffic rules, vulnerable road users may appear in unusual locations, and camera inputs may be degraded by weather, lighting, sensor faults, or occlusion.

"A model that performs well on clean images does not necessarily remain reliable when the road scene is unstructured or visually degraded," says Yuan Zheng, a researcher at Jiangsu Province Collaborative Innovation Center of Modern Urban Traffic Technologies, Southeast University. "Our goal was to build a test framework that makes these weaknesses visible under controlled and reproducible conditions."

USB reorganizes driving samples into synchronized six-camera inputs and expands each scene with 20 realistic corruption types, including backlight, glare, darkness, fog, rain, snow, sandstorm, smoke, motion blur, zoom blur, compression, quantization, bit errors, frame loss, lens obstruction, splash, and sensor failure. The benchmark then asks driving-related questions covering vulnerable road user perception, object motion prediction, traffic light recognition, environment description, causal explanation, and three-second ego-vehicle planning.

Unlike evaluations that rely only on simple accuracy, USB combines several scoring strategies. Multiple-choice questions are scored by exact matching and aggregated with expert-derived importance weights from 55 autonomous driving professionals. Free-text answers are evaluated with a GPT-4o-based scoring rubric that examines whether responses identify critical factors, remain consistent with visual evidence, explain causal relationships, and describe the road environment clearly.

Seven open-source VLMs with fewer than 8 billion parameters were tested: Qwen2.5-VL-7B-Instruct, Qwen3-VL-8B-Instruct, InternVL3.5-8B-Instruct, Keye-VL-1.5-8B, Phi-3.5-Vision-Instruct, LLaVA-v1.6-Mistral-7B, and LLaVA-v1.5-7B. Qwen3-VL achieved the highest unweighted multiple-choice accuracy at 0.6771 and the best causal explanation score at 8.24. InternVL3.5 achieved the highest weighted multiple-choice accuracy at 0.6660, while Qwen2.5-VL achieved the best road and environment description score at 9.44.

The results show that current VLMs still struggle with higher-level driving reasoning. Perception-related questions were generally easier, while object motion prediction and especially three-second ego planning were much harder. In the three-second planning task, all models scored low, indicating that recognizing scene elements does not automatically translate into reliable planning decisions.

The study also found that model rankings changed across corruption types. Some models remained strong under certain visual degradations but dropped sharply under others. More importantly, several models could still generate fluent and plausible answers when visual evidence was severely degraded or even absent. Under sensor failure, for example, some scores improved rather than declined, suggesting that models may sometimes rely on answer priors or general driving knowledge instead of the actual image content.

"High scores under missing or corrupted visual input should not be mistaken for genuine scene understanding," says Shen Li, a researcher at Department of Civil Engineering, Tsinghua University. "They can instead indicate weak visual grounding, which is a critical issue for safety-related applications."

The researchers argue that autonomous driving VLMs need evaluation settings that go beyond clean-image accuracy and include unstructured environments, visual corruption, task importance, and free-text reasoning quality. USB provides reproducible baselines and a diagnostic framework for identifying when models perceive, reason, or plan unreliably under challenging road conditions.

About Communications in Transportation Research

Journal

Communications in Transportation Research

DOI

10.26599/COMMTR.2026.9640034

Article Title

Unstructured Scene Benchmark (USB): Which VLM Performs Better in Autonomous Driving?

New method to create customizable virtual road networks for autonomous driving testing

Tsinghua University Press

Overview of road network generation — image:
Overview of the proposed complexity-controllable road network generation framework. The workflow is organized into three systematic stages: road element modeling extracts fundamental road topologies elements from real roads, establishing parameterized models with quantitative complexity labels;
road network synthesis employs heuristic optimization algorithms to interconnect element instances, balancing multiple core objectives including network fidelity and scenario density while enforcing non-intersection and non-overlapping constraints to generate topologically valid, seamlessly connected road networks; scenario map construction converts the abstract topological network into a high-fidelity high-definition (HD) map compliant with the OpenDRIVE standard, incorporating detailed geometric features such as lane tapers, curve transitions, and standardized lane markings.
view more
Credit: Communications in Transportation Research

Researchers have developed a method to generate customizable virtual road networks with adjustable complexity, which cuts required test mileage by over 30% while preserving real-world road features. This enables faster, more efficient identification of functional defects and performance boundaries in autonomous vehicles.

The team published their study in Communications in Transportation Research (https://doi.org/10.26599/COMMTR.2026.9640024).

The "Mileage Compression" Breakthrough

Traditional testing relies on driving millions of miles on real roads. This new method analyzes road complexity and creates a "greatest hits" version of a city's streets.

Researchers utilized cartographic data from Xi'an to validate their model. They found that by intelligently assembling road elements—such as merge, diverge and complex intersections—the generated networks achieved a 32.4% compression in test mileage.

“The infeasibility of on-road testing that requires millions of kilometers has been a proven bottleneck. Our method integrates real-world cartographic data with optimization algorithms to generate networks that preserve real-world road features while eliminating redundant mileage.” says Yu Zhu, a researcher at the School of Information Engineering at Chang’an University.

Pinpointing Safety Limits
The research didn't just save time; it proved effective at finding flaws. When tested against a leading autonomous driving algorithm, the virtual networks successfully identified functional defects that might be missed in standard testing.

Key findings include:

High Fidelity: The virtual roads have a geometric error of less than 1.5% compared to real Xi'an roads.
Risk Correlation: The "complexity" of the generated roads showed a strong correlation (r > 0.84) with driving difficulty, meaning harder roads lead to more rigorous testing.
Defect Detection: In simulation testing, test scenarios constructed using the generated road networks successfully exposed the performance limits of the autonomous driving algorithm under lateral approach conditions, where the test vehicle failed to reliably avoid collisions even after correctly detecting the approaching target vehicle.

“The experiment results demonstrated that the test scenarios constructed using generated networks could successfully capture both the collision avoidance performance limits of the algorithm and its functional defects under lateral approach conditions.” explains Professor Zhigang Xu, the project’s supervisor.

Methodology and Future Outlook

The team utilized a combination of Particle Swarm Optimization (PSO) and Simulated Annealing (SA) algorithms to stitch together road segments without overlaps. This "non-intersecting" assembly allows for continuous scenario testing, where a car drives through a series of complex maneuvers without stopping.

While the current model focuses on standard intersections, the authors note future work will expand the library to include roundabouts and highway interchanges

About Communications in Transportation Research

Journal

Communications in Transportation Research

DOI

10.26599/COMMTR.2026.9640024

Article Title

Complexity Controllable Road Network Generation for Virtual Testing of Autonomous Driving

How to accurately predict the future trajectories of mixed-flow vehicle platoons in traffic oscillation scenarios?

Tsinghua University PressLinkedIn

The framework of the CPIDLM model — image:
The CPIDLM network architecture is illustrated in Figure. The architecture consists of four modules labeled (a) through (d). (a) The driver heterogeneity and uncertainty feature learning module extracts causal behavioral features based on a causal graph of driver behavior, enabling effective modeling of driver heterogeneity and uncertainty. This module explicitly expresses these variations through five IDM parameters that vary with traffic conditions and time. The parameters output to characterize driver heterogeneity can be directly applied to vehicle platoon dynamics modeling and simulation. (b) The physical module uses these five parameters to compute physical predictions of driver behavior according to the IDM’s computational graph. (c) The data-driven vehicle motion feature learning module builds vehicle following behavior characteristics solely from large-scale vehicle platoon data to predict vehicle trajectories purely from data. (d) The adaptive weight module dynamically outputs gating thresholds based on each vehicle’s state to integrate the physical predictions with those from the data-driven module. The key advantage of this framework is that the predictions are constrained by physical prior formulas, ensuring consistency with physical realities and producing more reliable and stable results.
view more
Credit: Communications in Transportation Research

To answer this question: How to accurately predict the future trajectories of mixed-flow vehicle platoons in traffic oscillation scenarios? This study introduces CPIDLM, a hybrid AI model that significantly improves vehicle movement predictions during traffic speed changes by combining physical driving laws with individual human behavior patterns. By explicitly accounting for driver uncertainty and how vehicles affect one another, this approach reduced distance prediction errors by 16.7% compared to existing methods, providing a more reliable foundation for autonomous vehicles.

The team published their study in Communications in Transportation Research (https://doi.org/10.26599/COMMTR.2026.9640029).

Researchers at Central South University and Monash University have developed a cutting-edge artificial intelligence framework designed to solve one of the most frustrating challenges in modern transportation: the stop-and-go traffic oscillation. The new approach, known as the Causal Physics-Informed Deep Learning Model (CPIDLM), allows autonomous vehicles to see and predict the erratic movements of human drivers ahead of them with unprecedented accuracy.

Cracking the Code of Human Driving Uncertainty

Traditional autonomous driving systems often struggle in mixed traffic because human driving behavior is inherently heterogeneous and uncertain. Different drivers have varying reaction times and safety distances, which causes speed fluctuations to ripple through a line of cars like a wave.

“Autonomous vehicles relying exclusively on local perception remain susceptible to traffic oscillations, primarily due to the uncertainty of human driving behavior,” explains Jipu Li, a researcher at the School of Traffic and Transportation Engineering at Central South University. These oscillations increase safety risks, diminish comfort, and spike energy consumption.

A Hybrid Approach: Merging Physics with AI

To tackle this, the research group moved beyond standard black box AI. Their CPIDLM framework uses two clever mechanisms to stabilize predictions:

Causal Discovery: The model uses a Causal Graph Attention mechanism to figure out exactly how one vehicle’s actions cause a reaction in the next, effectively disentangling complex traffic interactions.

Physics Integration: Instead of guessing blindly, the AI is constrained by the laws of traffic physics (specifically the Intelligent Driver Model), ensuring its predictions never violate real-world kinematic limits.

Adaptive Weighting: A dynamic gating module decides in real-time whether to trust the physical laws or the data-driven patterns more, depending on the current traffic state.

Significant Gains in Road Safety and Efficiency

The results of the study, validated against real-world trajectory datasets like NGSIM, were striking. The CPIDLM reduced gap prediction errors by 16.7% compared to existing state-of-the-art models.

“The results show that integrating physical prior models reduces the deep learning models’ reliance on pure data,” says Ye Li, a Professor at Central South University. “This establishes a powerful new paradigm for helping intelligent connected vehicles make optimal decisions even when surrounded by unpredictable human drivers”.

The researchers hope this work will lay the foundation for more stable and energy-efficient autonomous fleets, preventing the safety risks typically associated with sudden highway braking.

About Communications in Transportation Research

Journal

Communications in Transportation Research

DOI

10.26599/COMMTR.2026.9640029

Article Title

Vehicle Platoon Trajectory Prediction Under Traffic Oscillation: A Causal Physics-Informed Deep Learning Approach

Can China VI diesel trucks really cut emissions on real roads?

Tsinghua University Press

image:
Operating mode-specific PN emission factors from different vehicle groups
view more
Credit: Communications in Transportation Research

Researchers from Shenzhen Polytechnic University, the University of California, Riverside, and Tsinghua University evaluated real-world emissions from China V and China VI heavy-duty diesel trucks using portable emissions measurement systems (PEMS). The study found that compliant China VI trucks can significantly reduce NOx and particle number (PN) emissions, while some vehicles still showed excessive emissions under real driving conditions, highlighting the need for stronger in-use supervision.

The team published their study in Communications in Transportation Research (https://doi.org/10.26599/COMMTR.2026.9640033).

“We wanted to understand whether the stricter China VI regulations can truly deliver cleaner emissions under actual driving conditions,” says Dr. Weixia Li from Shenzhen Polytechnic University. “Real-road testing is critical because vehicle performance in laboratories may differ substantially from real-world operations.”

Real-road testing reveals mixed emission performance

Heavy-duty diesel trucks are major contributors to nitrogen oxides (NOx) and particulate matter (PM) emissions in the transportation sector. To reduce these pollutants, China implemented the China VI heavy-duty vehicle emission standard in 2021, introducing stricter NOx limits, particle number (PN) regulations, and mandatory portable emissions measurement systems (PEMS) testing.

In the study, the researchers conducted real-road emission tests on 24 heavy-duty diesel trucks, including both China V and China VI vehicles from different manufacturers. The team measured NOx and PN emissions under various driving conditions.

The results showed that compliant China VI vehicles reduced PN and NOx emissions by 51.5% and 59.6%, respectively, compared with China V vehicles. However, some China VI trucks still produced unexpectedly high emissions under real-road conditions. Among the tested vehicles, 28.6% exceeded PN limits and 23.8% exceeded NOx limits.

“Some non-compliant China VI vehicles even emitted more pollutants than older China V trucks,” says Dr. Li. “This suggests that stricter standards alone are not enough without effective in-use monitoring and maintenance.”

SCR and EGR systems play important roles

The study also examined how after-treatment systems performed under different operating conditions. The researchers found that SCR systems achieved NOx reduction efficiencies ranging from 76.6% to 96.8%, although performance decreased during low-speed driving possibly because of lower exhaust temperatures. In addition, vehicles equipped with EGR systems generally showed lower NOx emissions across most driving conditions, indicating that EGR can effectively complement SCR-based emission control strategies.

“Our findings highlight the importance of maintaining stable after-treatment performance under complex driving conditions,” says Dr. Li. “Future regulations should place greater emphasis on real-driving compliance and low-speed emission control.” The researchers believe the study can support future emission regulations, cleaner freight transportation systems, and more effective real-road emission supervision.

About Communications in Transportation Research

Journal

Communications in Transportation Research

DOI

10.26599/COMMTR.2026.9640033

Article Title

On-road Evaluation of Emission Control in China V-VI Heavy-duty Diesel Trucks

Tuesday, October 13, 2026

Bacteria DNA has evolved to withstand copper

Copper-based coatings are often used to slow bacterial growth on marine surfaces — but bacteria have adapted

American Institute of Physics

Coated and uncoated PVC panels placed in the Chesapeake — image:
Coated and uncoated PVC panels (top), which encountered bacterial growth after being placed in the Chesapeake Bay (bottom, left). The bacteria were then assayed and their genetics studied (bottom, center and right).
view more
Credit: Kenan P. Fears

WASHINGTON, Oct. 13, 2026 — Marine bacteria have evolved to bypass the copper-based coatings placed under ships to stop the growth of barnacles, algae, and other marine life. This necessitates new methods for managing biofouling — the unwanted accumulation of marine organisms — on ship hulls.

In Biointerphases, an AVS journal published by AIP Publishing, researchers from the U.S. Naval Research Laboratory and Duke University studied the genetic characteristics of various bacteria species responsible for their survival on copper surfaces, a step toward inhibiting these copper-resistant genes.

“We have observed increasing copper resistance in some of the ports our team frequents over the past decade, particularly in ports with heavy tourist [or] industry shipping activity,” said author Kenan Fears. “We wanted to identify the resistance pathways across a diverse microbial community to provide insight into how members of a complex community could be acting symbiotically to resist toxic environments.”

The researchers placed pieces of PVC panels in the Chesapeake Bay, some coated with a copper-based coating and some left uncoated. After two weeks, they extracted the DNA and sequenced the genomes of the bacterial colonies that grew on the panels to identify the copper-resistant genes and tested each colony’s tolerance against copper-based water treatments.

Surprisingly, the bacteria species had a lot in common regardless of whether they grew near the copper-based coating. The bacteria from the control colonies — the ones that did not initially grow on the copper-coated surfaces — were just as copper-resistant and could thrive at copper concentrations above 70 milligrams per liter.

“To put that in perspective, the U.S. Environmental Protection Agency limit in seawater for a short-term spike in copper is 18 micrograms per liter,” Fears said.

This is a likely marker of evolution and bacteria adapting to their new environments.

“Copper was not always in a water-soluble state; it only became so after oxygen accumulated in Earth’s atmosphere,” Fears said. “However, as environmental copper loading has increased, the bacteria with the best defense mechanisms continued to reproduce.”

The researchers hope to use their findings to target the molecular pathways that help bacteria express copper-resistant genes, weakening their ability to survive in copper-rich environments.

“Our research suggests the smarter path moving forward is to design antifouling coatings that select microbial communities that either help deter macrofouler settlement — like barnacles or tubeworms — or do not impede the release of biocidal compounds,” Fears said. “With increasing regulatory pressure on copper use in antifouling coatings, there is an urgent need for alternatives.”

###

The article “Extreme copper tolerance and genomic adaptation in early-stage marine biofilms” is authored by Kailey N. Richard, Sara M. Tuck, Sophie M. Colston, Maryssa A. Beasley, W. Judson Hervey, Kevin Le, Sean Brown, Katherine J. Franz, and Kenan P. Fears. It will appear in Biointerphases on Oct. 13, 2026 (DOI: 10.1116/6.0005640). After that date, it can be accessed at https://doi.org/10.1116/6.0005640.

ABOUT THE JOURNAL

Biointerphases, an AVS journal published by AIP Publishing, emphasizes quantitative characterization of biomaterials and biological interfaces. As an interdisciplinary journal, a strong foundation of chemistry, physics, biology, engineering, theory, and/or modelling is incorporated into originated articles, reviews, and opinionated essays. See https://pubs.aip.org/avs/bip.

ABOUT AVS

AVS is an interdisciplinary, professional society with some 4,500 members worldwide. Founded in 1953, AVS hosts local and international meetings, publishes five journals, serves members through awards, training and career services programs, and supports networking among academic, industrial, government, and consulting professionals. Its members come from across the fields of chemistry, physics, biology, mathematics, engineering, and business and share a common interest in basic science, technology development and commercialization related to materials, interfaces, and processing.

###

Journal

Biointerphases

DOI

10.1116/6.0005640

Article Title

Extreme copper tolerance and genomic adaptation in early-stage marine biofilms

Article Publication Date

13-Oct-2026

Newly discovered antiviral system in bacteria is ‘strangest of its kind’

The complex Rube Goldberg-like molecular machinery could inspire novel technology to study and treat disease

Gladstone Institutes

Gladstone scientist Karen Zhang — image:
Gladstone scientists, including Karen Zhang, uncovered a counterintuitive class of bacterial defense systems, known as type VI retrons, that can be adapted to engineer more efficient gene editing tools.
view more
Credit: Michael Short/Gladstone Institutes

SAN FRANCISCO—Researchers at Gladstone Institutes have uncovered an extraordinarily unusual bacterial immune system that operates in reverse of known defenses.

The discovery, published in Molecular Cell, details the mechanics of "type VI retrons"—a uniquely counterintuitive class of bacterial defenses. More than just a biological curiosity, decoding these complex systems provides a molecular blueprint that scientists can adapt to engineer more efficient DNA-editing tools for human health and therapies.

“This is, by far, the strangest of the retron systems that we or anyone else has come across,” says Gladstone Investigator Seth Shipman, PhD, senior author of the new study. “These retrons operate with Rube Goldberg-level complexity, using molecular mechanisms that we would never have thought to invent, but that we’ll use in future technologies.”

Bacteria have long evolved specialized defenses to destroy phages, which are viruses that target bacterial cells. A particularly powerful line of defense, known as retrons, recognize viral attacks and trigger self-destruction to protect the broader bacterial colony.

Because retrons can also efficiently modify and deliver single-stranded DNA, researchers including Shipman and his team are working to harness them to rewrite disease-causing mutations inside human cells with unprecedented precision.

Inverted Expectations

Many types of retron systems exist, but they all follow the same basic principle: they sense when a specific strain of phage has attacked the bacterial cell, and they then unleash a toxic protein to kill the cell. This self-destruction stops the phage before it has a chance to multiply and spread to neighboring cells in the same colony.

Decades of research have led to the general understanding that all retrons produce a toxin that, in uninfected cells, is constantly present but deliberately repressed by small, single-stranded DNA molecules. When a phage invades, it perturbs the DNA and causes it to release the toxin.

“The way we typically detect retrons inside uninfected cells is by looking for this single-stranded DNA,” says Karen Zhang, a PhD student in Shipman’s lab and lead author of the study. “But type VI retrons caught our attention because they didn’t appear to have this signature. It was very odd, and we wondered if they might be broken.”

But when the scientists tested certain bacteria containing type VI retrons, they found that the retrons were still able to successfully defend their cells against phages.

“That kicked off a long molecular odyssey,” Shipman says. “Whenever we tried the experiments we normally use to study retrons, they would fail, so we kept having to think of new ways to figure out how these retrons work.”

The experimental saga ultimately revealed that, unlike other retrons, type VI retrons do not produce noticeable levels of single-stranded DNA nor toxin in uninfected cells. Instead, the entry of a phage into a cell is needed to trigger the retron to start producing single-stranded DNA.

Once it builds up, the DNA then interacts with an RNA that has been blocking production of the toxin. This interaction releases the block, allowing toxin production to ramp up.

The toxin appears to be a tiny protein that embeds itself into the membrane surrounding the cell, which destabilizes the cells and kills it.

“It’s almost the mirror-image version of how we traditionally view retrons, because the DNA switches on toxin production instead of inhibiting an already-produced toxin,” Zhang says. “It would be fascinating to explore how bacterial evolution could have resulted in such a complex molecular system—whether type VI retrons are older or newer than other retrons, or whether they’re a stepping stone between different types.”

Rethinking Retron Research

To decipher the unique machinery of type VI retrons, the team had to concoct novel experiments and repurpose existing methods in new ways. One early breakthrough, for instance, arose from the repeated, mysterious failure of standard methods to detect DNA made by type VI retrons in uninfected cells.

The researchers turned to a different tool called Spacer-seq previously developed in Shipman’s lab. This tool harnesses the molecular components of another kind of bacterial immune system, CRISPR-Cas, to create a record of gene activity inside a cell.

“We had just about given up hope that any DNA was being made when this approach finally revealed a very, very low level of production in uninfected cells,” Zhang says.

The finding motivated the team to keep digging, and further experiments showed that the DNA only ramps up to defensive levels after the phage enters.

Nature-Inspired Tools and Treatments

The unique machinery of type VI retrons is already inspiring the development of novel tools in Shipman’s lab, including new synthetic sensors and switches that can control gene activity.

“This system represents a unique way to potentially build a molecular sensor and switch encoded in DNA and RNA, that no one has thought of before,” Zhang says. “Synthetic sensors and switches have many possible applications, both in treating disease and in designing experiments to better understand how certain diseases operate.”

The tiny, highly potent toxins produced by the type VI retrons could also prove useful in a medical context. The scientists are working on ways to incorporate these toxins into new types of antimicrobial therapies, which could target and kill bacteria that cause human disease.

“This study highlights why it’s so important for a molecular technology development lab like ours to keep looking to nature for inspiration,” Shipman says. “We can use what evolution has given us as starting points to build complex machines that would have been tremendously hard to even conceptualize.”

###

About the Study

The study, “Reverse transcribed ssDNA derepresses translation of a retron antiviral protein,” was published in Molecular Cell on October 13, 2026. In addition to Shipman and Zhang, authors are Matías Rojas-Montero, Darshini Poola, and Alejandro González-Delgado of Gladstone; Josepha Klas, Arturo Carabias, Dennis Zhang, Mario Mestre, Ruiliang Zhao, Ana Dávila-Hidalgo, Guillermo Montoya, and Rafael Pinilla-Redondo of the University of Copenhagen; and Lieza Chan and Kliment Verba of UC San Francisco.

The work was supported by the National Science Foundation, the Robert J. Kleeberg Jr. and Helen C. Kleberg Foundation, the Gordon and Betty Moore Foundation, the Lundbeck Foundation, the Fulbright U.S. Student Program, the Junta de Castilla y León and European Social Fund Plus, the European Molecular Biology Organization, the Novo Nordisk Foundation, the European Research Council, and the Villum Foundation.

About Gladstone Institutes

Gladstone Institutes is an independent, nonprofit life science research organization that uses visionary science and technology to overcome disease. Established in 1979, it is located in the epicenter of biomedical and technological innovation, in the Mission Bay neighborhood of San Francisco. Gladstone has created a research model that disrupts how science is done, funds big ideas, and attracts the brightest minds.

Journal

Molecular Cell

DOI

10.1016/j.molcel.2026.09.026

Article Title

Reverse transcribed ssDNA derepresses translation of a retron antiviral protein

Article Publication Date

13-Oct-2026