Gradient-based trajectory optimization often encodes safety specifications in Signal Temporal Logic (STL) through a smooth approximation of robustness, a quantitative measure of how well a trajectory satisfies the spec. For safety-critical systems, that approximation must be sound—nonnegative only on trajectories that actually satisfy the specification—and it must keep allocating gradient weight to the samples that violate the specification, even when the time horizon is long. The authors note that common choices, including log-sum-exp and its sound variant, fail at least one of these requirements.
To fix this, the paper introduces lower approximations of the min and max operators whose gradients are sparsemax projections. The resulting robustness lies below the exact value, which guarantees soundness, and its error is bounded independently of the number of samples. Moreover, the gradient is exactly zero only on samples beyond a data-dependent threshold, so the weight on the worst violating samples has a positive lower bound at every horizon.
In experiments with a planar unicycle and a seven-joint manipulator operating beside a person, the log-sum-exp approximation reported positive robustness on a trajectory that actually violated the specification. As the manipulator's horizon grew from 10 to 80 seconds, the gradient weight of the sound variant on violating samples dropped from 0.22 to 0.04, while the sparsemax weight remained at 1. In constrained unicycle optimization, the log-sum-exp constraint returned a violating trajectory, whereas the sparsemax constraint produced one that satisfied the specification with the required margin.