The research, conducted by Quanda Zu and Shaobo Cai of Wenzhou Polytechnic University together with Jianwei Li of Jiangsu University, tackles a stubborn weakness in how autonomous cars have traditionally judged safety. Conventional systems lean on threshold rules built around surrogate safety measures such as time-to-collision, or TTC, the time remaining before two vehicles collide if neither changes speed. TTC is intuitive, but it is a blunt instrument. It assumes a constant relative velocity, ignores lateral motion entirely, and returns an infinite value when two cars cruise at the same speed, no matter how terrifyingly close they are. Two situations with identical TTC values can demand wildly different braking responses, because the underlying accelerations, gaps and speeds differ.

The team’s answer is a nine-rule Takagi-Sugeno fuzzy engine trained through the ANFIS hybrid-learning procedure, in which membership functions are tuned by gradient descent while rule consequents are estimated by recursive least squares. The model takes two inputs: the normalized critical deceleration demand, essentially how hard the ego vehicle would have to brake to match its neighbor’s speed within the available gap, and the normalized minimum gap expressed as a time headway. Nine overlapping linguistic rules, spanning states from high deceleration with low headway down to non-braking conditions with generous margins, produce a continuous risk index between zero and one. Because the membership functions deliberately overlap, small changes in the traffic picture cause smooth, graded shifts in risk rather than the abrupt binary flips that plague threshold logic and produce so-called decision chattering.

That risk index does not act alone. It feeds a two-phase decision mechanism that treats a lane change not as a single event but as a spatially structured maneuver. The lateral path is split at the midpoint of the lane width: before that boundary, the maneuver is reversible, and the car can abort and generate a smooth return trajectory; beyond it, the vehicle is committed, and the safest response is usually controlled deceleration within the target lane rather than a destabilizing swerve back. The system evaluates four surrounding vehicles simultaneously, weighting the target-lane lead vehicle most heavily at 0.40, followed by the target-lane follower at 0.30, the origin-lane follower at 0.20, and the origin-lane leader at 0.10. Crucially, an asymmetric veto rule prevents averaging from hiding a single catastrophic interaction, so one imminent collision risk cannot be diluted by three calm neighbors.

For SAE Level 3 operation, the framework adds a takeover management layer that models handover as a staged process rather than an instant. It budgets three distinct intervals: driver reaction time, physical control resumption time, detected when steering torque exceeds two newton-meters or brake pedal displacement passes five percent, and vehicle stabilization time. If the available time budget before reaching a hazard falls short of the total takeover time, the system escalates the warning and simultaneously commands whatever autonomous mitigation remains available, from gentle braking to a minimum-risk maneuver. The design explicitly prioritizes safety over maneuver completion; if a lane change cannot be finished safely within a predefined window near a highway exit, the car decelerates and waits.

When hazards do appear mid-maneuver, the trajectory planner responds with quintic Bezier curves, six control points arranged so that lateral velocity and acceleration vanish at both ends of the path. This imposes second-order continuity, meaning no jerky curvature discontinuities during an abort. Before the commitment boundary, the planner reshapes or cancels the nominal path using the current vehicle state as a new anchor; after it, the geometry is preserved and the longitudinal velocity profile absorbs the disturbance instead. A lateral LQR controller with feedforward curvature compensation tracks these paths, and a fuzzy brake-pressure layer translates the aggregated risk into graded braking commands, keeping actuation as interpretable as the reasoning behind it.

The quantitative results are striking. Benchmarked against four classical baselines, a static Mamdani fuzzy system, a TTC trigger at two seconds, a deceleration-demand trigger, and a constant time-headway policy, the ANFIS framework doubled the warning lead time, from 0.86 seconds for TTC to 1.84 seconds, giving the vehicle nearly an extra second to act. The minimum two-dimensional clearance between vehicles grew from 2.08 to 4.82 meters, while the maximum deceleration demand fell from a bone-jarring 4.92 to a comfortable 2.34 meters per second squared. Threshold-based competitors exhibited three to five mode switches per maneuver near their decision boundaries, a signature of chattering induced by sensor noise, whereas the proposed system recorded zero, with lateral jerk held to 0.82 meters per second cubed.

Robustness testing is where the study stakes its boldest claim. Across 1,000 Monte Carlo simulations with randomized speeds between 90 and 120 kilometers per hour, sensor noise, road friction coefficients spanning wet asphalt to dry concrete, and cut-in timing variations, the proposed framework produced zero collisions, a 0.30 percent near-miss rate and a 1.20 percent false-alarm rate. The TTC baseline, by contrast, collided in 2.90 percent of runs under low-friction late-cut-in conditions, precisely because fixed thresholds cannot account for reduced tire-road adhesion. The multi-input fuzzy reasoning, combined with Schmitt-trigger hysteresis, also acted as a low-pass filter on noisy measurements, suppressing the false alarms that plagued single-variable triggers, which misfired more than eight percent of the time in high-noise environments. Meanwhile, the mean computational execution time was just 1.42 milliseconds, comfortably inside the 10-millisecond budget of a 100-hertz control cycle, and the trained model achieved a coefficient of determination of 0.9815 with a root-mean-square error of 0.0248 on validation data.

The deeper significance of the work lies in its stance on the interpretability-versus-capability trade-off. Deep reinforcement learning policies can extract subtle features from dense traffic, but their black-box neural topologies offer no traceable diagnostic reasoning, which is a serious obstacle to automotive certification under standards like ISO 21448, the safety-of-the-intended-functionality framework. The nine fuzzy rules here remain readable: a human engineer can inspect exactly why the system aborted a maneuver or issued a takeover request, and the learned parameters adjust those rules without obscuring their logic. In that sense, the study argues that adaptivity and transparency need not be enemies, and that a neuro-fuzzy middle path can deliver both learned precision and certifiable structure.

The authors are careful about scope. Everything demonstrated so far comes from high-fidelity co-simulation, pairing SCANeR Studio vehicle dynamics and sensor models with MATLAB/Simulink control code on a modeled Audi A8 platform, with the ego vehicle equipped with realistic radar and camera models while surrounding traffic followed scripted trajectories. Hardware-in-the-loop rigs and real test-track trials remain necessary before any claim of real-world performance, and future extensions target dense urban environments, vulnerable road users, and vehicle-to-everything communication to share braking intentions among neighboring cars. Still, the headline finding stands on its own: by fusing braking demand and gap into one continuous, trainable, human-readable risk number, the framework gave a simulated Level 3 car nearly a full extra second of foresight, more than twice the clearance, and half the emergency braking effort, all while keeping every one of a thousand randomized highway encounters collision-free.

Subject of Research: Interpretable neuro-fuzzy risk assessment for lane-change decisions in Level-3 autonomous driving

Article Title: Adaptive neuro-fuzzy risk index for safe and interpretable lane-change decisions in Level-3 autonomous driving

Article References: Zu, Q., Cai, S., & Li, J. (2026). Adaptive neuro-fuzzy risk index for safe and interpretable lane-change decisions in Level-3 autonomous driving. Frontiers in Neurorobotics, 20, Article 1918469. https://doi.org/10.3389/fnbot.2026.1918469

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1918469

Keywords: autonomous driving, lane change, ANFIS, fuzzy logic, risk assessment, time-to-collision, SAE Level 3, takeover request, trajectory planning, Monte Carlo simulation, surrogate safety measures, vehicle safety