A new reinforcement learning system lets autonomous heavy trucks juggle safety, fuel costs, and schedule pressure instead of baking all three into one hidden number.
The framework, outlined in a newly published paper, trains policies using Generalized Policy Improvement and Proximal Policy Optimization to approximate a Pareto front across three competing goals: safety (measured by collisions and successful trips), energy efficiency (energy cost), and time efficiency (driver cost). The researchers tested it in a highway traffic simulation built for tactical decisions in heavy-duty trucks, running it across multiple traffic densities and comparing results against an analytical baseline. The resulting policies produced smooth, interpretable trade-off curves rather than a single fixed behavior. Crucially, the same trained model can shift its priorities on the fly, without retraining, by adjusting an input preference.
That flexibility matters because most self-driving stacks compress safety, speed, and cost into a single scalar reward, which quietly papers over the compromises being made. A fleet operator weighing fuel bills against delivery deadlines, or a regulator asking how a system handles risk, gets a more honest answer from a visible trade-off curve than from a black-box score. For commercial trucking, where margins and safety rules both bite, that kind of dial-able behavior could be more useful than a marginally smarter single policy.
Still, this is a simulation result, not a truck on I-80. A clean Pareto front doesn't tell you who decides where on that curve to sit, and that choice, not the algorithm, is where the real liability questions live.