Explainable AI traces how eVTOL drones dodge obstacles in simulated cities

Electric air taxis promise to thread their way through dense city skies, but the artificial intelligence that would steer them has a stubborn problem: nobody can see inside it. A research team now reports a simulation-based method that opens up the black box of reinforcement learning obstacle avoidance for electric vertical take-off and landing aircraft, mapping every steering decision back to physical flight variables a human engineer can read. The work, published in Discover Computing, combines a physics-guided learning controller with a feature-attribution pipeline and evaluates the resulting explanations against four quantitative criteria drawn from emerging aviation guidance on machine learning explainability.

The study addresses a tension that sits at the heart of advanced air mobility. Deep reinforcement learning has shown it can learn path planning under complex dynamic constraints through trial-and-error interaction between an agent and its environment, outperforming classical planners such as A*, the Vector Field Histogram method, and the Dynamic Window Approach, which depend on rapid mapping, predefined heuristics, and computations that grow exponentially with scenario scale. Yet the very data-driven flexibility that makes these policies effective also makes them opaque. Aviation regulations such as 14 CFR 25.1309 and its Chinese counterpart CCAR 25.1309 require system design requirements to be clear and traceable, so certification authorities care not only whether an aircraft avoided the obstacle but why a particular path was chosen. The authors argue that current certification frameworks, built around human-piloted logic and deterministic control, struggle to extract that kind of traceable evidence from black-box learning systems.

Their answer begins with the controller itself. The team built an obstacle avoidance strategy around the Soft Actor-Critic algorithm, a reinforcement learning method whose maximum entropy objective encourages efficient exploration in continuous action spaces, and wove the Artificial Potential Field method into its reward structure. Rather than using potential fields as a rigid controller, which conventional implementations do with known weaknesses such as susceptibility to local minima, the researchers treated them as a physics-informed inductive bias. The total potential field force decomposes into an attractive component pulling the aircraft toward its goal, a static repulsive component activated when the eVTOL enters a defined influence range around fixed obstacles, and a dynamic repulsive component that handles moving threats. For dynamic obstacles, the formulation introduces an obstacle implicit function, a normal repulsive matrix, and a tangential matrix that provides lateral guidance around the obstacle, combined into a perturbation matrix with velocity compensation based on the obstacle’s own motion. This design lets the force scale non-linearly as the aircraft closes in on a moving hazard.

The state space feeding the learning agent is deliberately semantic. Nine variables describe the aircraft’s situation: three relative position components to the mission target, three to dynamic obstacles, and three to static obstacles. The action space consists of displacement increments along the x, y, and z axes for each decision step. Because every input maps directly to flight physics, the researchers could later attribute decisions to features that mean something to a pilot or an engineer. The reward function blends three parts: a sparse reward of plus or minus 200 for arrival or collision, a continuous reward based on normalized progress toward the goal, and an APF guidance reward that scores how well each action aligns with the direction of the potential field force using a cosine similarity measure. Ablation experiments underscored why this composite matters. With the guidance reward removed, the agent failed to learn effective avoidance and success rates stayed near zero; without the continuous reward, learning improved but remained unstable. The complete design reached a 100 percent success rate quickly and held it steady through training.

Training and testing unfolded in simulated three-dimensional urban air mobility environments spanning 10,000 by 10,000 meters with a 400-meter altitude ceiling, run for 1,000 episodes on hardware that included two NVIDIA RTX 4090 GPUs. After training, the team ran 100 independent simulation tests with 100 target points distributed along a circular path of 5,000-meter radius, mixing cylindrical static obstacles with dynamic threats following varied trajectories. A sensitivity analysis varied the APF weight coefficient from 0.1 to 0.9, running 100 trials per value. The results traced a clear trade-off: smaller coefficients weakened the guidance reward and lowered success rates, while larger ones pushed the aircraft to keep greater clearance from obstacles at the cost of longer, more conservative paths. A t-test between the extreme settings found statistically significant differences in both minimum obstacle distance and average path length. Success rates hit 100 percent for coefficients between 0.4 and 0.9, and the researchers identified the 0.4 to 0.6 range as a balanced operating band. In denser scenarios with three, six, or nine static obstacles, the controller reached the target in every tested case without collision or safety-margin violation, though path length and time steps grew moderately as clutter increased.

With the controller validated, the explainability pipeline took over. The researchers collected more than 27,000 state-action pairs from 100 complete trajectories and trained four candidate surrogate models, standard GBDT, XGBoost, LightGBM, and CatBoost, to mimic the learned policy. Because reinforcement learning outputs are stochastic and the relationships between obstacle features and actions are highly non-linear, a tree-based surrogate was needed to absorb that complexity before explanation could begin. Data splitting was done at the trajectory level to prevent leakage, with 70 percent of trajectories for training and cross-validation and 30 percent held out entirely for final fidelity evaluation. XGBoost won the comparison decisively. Its coefficient of determination reached 0.9971 for the x-axis, 0.9976 for the y-axis, and 0.9753 for the more complex z dimension, with an average fidelity score of 0.950. GBDT matched the accuracy but needed 164.67 seconds of training time, undermining its utility at scale.

The explanation layer uses SHAP, the SHapley Additive exPlanations method, applied through TreeExplainer, which exploits the internal structure of tree ensembles to compute each feature’s marginal contribution to a prediction efficiently. SHAP values decompose any single action output into a base value plus weighted contributions from all nine state variables, with positive values pushing the prediction up and negative values pulling it down. The team then scored the whole framework on four indicators aligned with explainability objectives discussed in guidance from the European Union Aviation Safety Agency and SAE International, which is developing ARP 6983 for Level 1 AI applications. Fidelity, measured across four unseen test scenarios with randomized start and target locations and dynamic threats including constant-velocity, sinusoidal, and circular motion, came in at 98.80, 99.06, 97.73, and 91.73 percent, with most predictions falling within plus or minus 20 percent relative error of the original policy’s outputs. Completeness was satisfied structurally: the framework covers all nine state features and provides local attributions at every time step of the flight.

Consistency, the third indicator, tested whether the explanations track physical risk. The assessment logic distinguishes an avoidance phase, when distance to an obstacle is shrinking within an activation threshold, from a navigation phase, when the aircraft is not in a high-risk state. A step counts as consistent if the relevant obstacle feature’s SHAP contribution rises above its global average during avoidance, or if the target vector’s contribution dominates during navigation. Scores across the four scenarios ranged from 83.33 to 95.35 percent, which the authors read as evidence that the decision policy has a repeatable physical basis rather than emerging from stochastic fluctuation. Feature attribution, the fourth indicator, was demonstrated in a local case study of an aerial encounter in which two eVTOLs approach from opposite directions while threading between static obstacles. At the 133rd decision step, the three most influential features were the x-axis distance to the target and the x and y-axis distances to the dynamic obstacle. The attribution showed the aircraft maintaining goal awareness while the horizontal proximity of the oncoming vehicle drove a rightward and upward maneuver, matching the climbing behavior visible in the simulation.

Global analysis reinforced the picture. Comparing feature contribution charts at weight coefficients of 0.4 and 0.6 revealed an internal shift in decision logic: goal-oriented features dominated x and y displacement decisions, contributing 71.4 percent at the lower setting, but their share fell to 62.0 percent as the coefficient rose, while obstacle-avoidance features gained ground. This quantitative shift mirrors the increased safety margins observed in the trajectory tests, tying the internal logic of the policy to its external behavior. The authors are careful about the limits of all this. The work is confined to simulation; it does not constitute formal certification evidence, industrial validation, flight testing, or assessment by aviation authorities, and the alignment assessment with EASA and SAE objectives is framed as preliminary, development-stage support rather than compliance. They also note that no persistent local-minima trapping appeared in their reported experiments, though the issue may still demand attention in more cluttered or narrow-passage environments. Future directions include explainable conflict resolution in multi-agent coordination and robustness testing under complex meteorological and operational uncertainty. For now, the contribution is a working template: a controller whose inputs are physical, a surrogate that mirrors it with high fidelity, and an attribution method that turns each avoidance maneuver into a readable chain of causes.

Subject of Research: Explainability analysis of reinforcement learning-based obstacle avoidance decision-making for eVTOL aircraft in simulated urban air mobility environments

Article Title: Research on explainability of obstacle avoidance decision-making for eVTOL based on SAC-APF

Article References: Dong, L., Song, W., Chen, X., Liu, J., & Zhao, C. (2026). Research on explainability of obstacle avoidance decision-making for eVTOL based on SAC-APF. Discover Computing, 29(1), Article 848. https://doi.org/10.1007/s10791-026-10400-w

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10400-w

Keywords: eVTOL, urban air mobility, reinforcement learning, explainable AI, SHAP, XGBoost, artificial potential field, Soft Actor-Critic, obstacle avoidance, path planning, airworthiness, simulation