How can autonomous driving control adapt to passenger motion-sickness preferences?

The controller learns three weights, not collision risk

The framework treats personalization as a weighting problem. A finite-horizon optimal control problem combines a passenger-dependent stage cost with a collision-risk term, and the passenger part is a weighted sum of speed, acceleration comfort, and motion-sickness costs. The preference vector w = [w_v, w_M, w_a] holds the weights for speed, motion sickness, and acceleration comfort, and these are the quantities that learning updates.

Collision risk is deliberately excluded from that adaptation. The weights lambda_p and lambda_s balance passenger preference against collision risk and satisfy lambda_p + lambda_s = 1, but lambda_s is held fixed so that learning changes only the relative importance of the passenger-dependent objectives. The collision-risk term itself is a soft penalty built from time to collision, which is why the architecture needs a separate supervisor rather than relying on the cost alone.

Motion sickness accumulates instead of tracking instant acceleration

The paper's central modeling claim is that motion sickness cannot, in general, be represented by instantaneous acceleration alone, because it depends on the accumulated interaction between vestibular and visual information. The passenger model therefore uses the Motion Illness Symptoms Classification (MISC) scale, where M = 0 means no symptoms and M = 10 means vomiting, and predicts M_k with an individualized model.

That model first generates sensory conflict using the group-averaged Subjective Vertical Conflict model and then accumulates the conflict through an individualized two-parameter accumulation model. The susceptibility parameters theta_K and theta_T are assumed known for each passenger, and the paper states that identifying them is outside its scope. This matters for reading the results: the simulated susceptibility levels are inputs to the evaluation, not something the framework estimates.

Emotion comparisons update the weights offline

Preference learning does not ask passengers to score a trip directly. Instead, the framework assumes that for two episodes the passenger prefers the one with the higher episode-mean emotion score, which converts a measured affective response into the binary comparisons that preference learning needs. Each closed-loop episode is summarized by a reward-feature vector of accumulated negative speed, motion-sickness, and acceleration costs.

The posterior over w is updated after every L new eligible episodes using Bayesian inference and Markov chain Monte Carlo sampling. Eligible pairs must contain at least one new episode and satisfy a minimum difference in reward features; they are ranked for informativeness, labeled by the emotion comparison, and ties are discarded. Weights are parameterized as w = softmax(eta) with independent standard-normal priors so that they stay positive and sum to one, and the mean normalized posterior sample becomes the preference vector for subsequent episodes. The paper also assumes the passenger's underlying preference stays unchanged during each learning experiment, which is consistent with a fixed-parameter Bayesian model.

Learned weights change speed and MISC together

The fixed-seed evaluation compares initial uniform weights with final learned weights, averaged over four evaluation episodes per scenario, and reports speed, mean MISC, RMS acceleration, and mean emotion as initial-to-final pairs. The learned weights produce distinct closed-loop behaviors rather than one uniform shift.

For Alice, the final controller raises mean speed in all three susceptibility settings, accompanied by higher mean MISC and RMS acceleration. In the low-susceptibility case, mean speed rises from 17.34 to 18.81 m/s while mean MISC rises from 0.244 to 0.388. For Bob-low, the controller lowers mean MISC from 0.254 to 0.229 while keeping mean speed approximately unchanged, producing a small emotion gain. Bob-high accepts higher MISC and acceleration in exchange for higher mean speed, with mean emotion rising from 13.781 to 14.089. Bob-medium moves the other way: speed, MISC, and RMS acceleration all decrease, but mean emotion falls from 15.846 to 15.095.

A four-episode mean can hide one bad episode

The paper is explicit that an improvement in the four-episode mean does not imply improvement in every evaluation episode. For Alice-low and Alice-medium, only one of the four evaluation seeds improves in emotion, but that increase is large enough to make the four-seed mean positive. For Bob-medium, a single evaluation seed loses roughly 4.3 in emotion, which outweighs the improvements in the other three seeds.

The clearest case is Zoey-high. Mean speed rises from 18.64 to 18.89 m/s, but one of the four evaluation episodes crosses the threshold M = 1 and reaches a maximum MISC of 1.079, staying at or above M = 1 for 7.75 s. That activates the large penalty and produces a mean emotion of -135.207 for that episode, pulling the four-seed mean from 0.746 to -33.250. The maximum MISC values in the other three final evaluation episodes are 0.833, 0.864, and 0.750. The authors read this as the difference between learning a smooth weighted surrogate objective and satisfying a threshold-sensitive passenger preference, since the learning mechanism does not impose M < 1 as a constraint.

No collisions observed, but not a safety guarantee

No physical collision is observed in the 2304 learning episodes or the 1188 evaluation episodes of the full method, and the safety supervisor intervenes on 1.621% of all learning control steps. For the final learned controllers, the mean intervention rate across the four evaluation episodes ranges from 0.625% to 4.271%. The supervisor reaches its best-effort branch, where no admissible candidate is found in the finite candidate set, on 34 learning control steps and 15 evaluation control steps.

The paper states directly that the absence of observed collisions should be read as an empirical result of these simulations and not as a formal guarantee of collision avoidance. Two details explain why. First, the logged time to collision is computed from the instantaneous relative velocity and collision ellipse at the pre-control state, while the planner evaluates risk along its predicted trajectory, so a reported minimum TTC can fall below the 5 s penalty threshold without violating a hard constraint, because the problem imposes no such constraint. Among the final-weight evaluations, Alice-high has the smallest mean episode-minimum TTC at 1.550 s. Second, the nominal vehicle-motion limits are exceeded on 86 learning control steps and 20 evaluation control steps; of these 106 occurrences, 90 arise during numerical planner fallback and 16 during safety-supervisor intervention. The largest recorded lateral acceleration magnitude is 11.706 m/s^2 against a nominal limit of 2.2 m/s^2, so satisfying the planning constraints on the nominal optimized trajectory does not guarantee satisfying them after a fallback or an override.

Acceleration and MISC features are not interchangeable

The feature ablation repeats the full 256-episode learning protocol and fixed-seed evaluation with two reduced preference models: V A keeps only speed and acceleration, and VM keeps only speed and MISC, with the two weights initialized uniformly at 0.5 in each reduced model. Removing the acceleration feature in VM raises aggregate RMS acceleration to 0.719 m/s^2, compared with 0.561 m/s^2 for the full model and 0.554 m/s^2 for V A, indicating that the explicit acceleration feature influences the smoothness of the resulting closed-loop trajectories.

Removing the MISC feature in V A produces two motion-sickness threshold crossings among the 12 Zoey evaluation episodes, compared with one crossing for the full model and none for VM. V A does avoid threshold crossing in all four Zoey-high evaluation episodes, so acceleration regulation can indirectly reduce motion-sickness accumulation in some realizations, but the paper does not treat acceleration as a consistent replacement for the explicit MISC feature. The ablation also undercuts a simple more-features-is-better reading: the full three-feature model achieves the highest four-seed mean emotion in only one of the nine scenarios, while V A and VM each do so in four. Because acceleration also influences subsequent motion sickness, the features are correlated over an episode, and different weight allocations can explain similar pairwise comparisons; the authors suggest more data, richer preference feedback, or additional structure in the preference model to improve identifiability.

Where the threshold case decides the reading

The most decision-relevant split in this paper is not between passengers but between scenarios where the learned weighted objective stays below the noticeable-symptom threshold and the one where it does not. Zoey-low and Zoey-medium gain both mean speed and mean emotion while the MISC state stays below the noticeable-symptom threshold in the evaluation episodes. Zoey-high gains mean speed yet fails on a single episode, and that single failure dominates the four-seed average.

That contrast sets the priority for anyone reading these results: the framework's value depends on whether a passenger's preference is smooth or threshold-like, because the current preference-learning mechanism does not impose M < 1 as a constraint and a weight vector that performs well over many episodes can still fail to prevent an individual threshold crossing. The paper's own next steps follow the same split, pointing to human-subject feedback, uncertainty in motion-sickness estimation, and preference models that better represent threshold-dependent and time-varying passenger responses. Until those are addressed, the reported gains should be read as simulation results for three passenger profiles at three susceptibility levels, not as a validated personalization guarantee.

Sources used in this answer