Most Robot Safety Methods Fail Under Real Uncertainty — pdSTL Finally Delivers Formal Guarantees

Autonomous robots operating in unpredictable environments face a fundamental challenge: satisfying complex temporal safety constraints despite stochastic dynamics and sensing noise. Existing Signal Temporal Logic (STL) methods offered robustness measures for gradient-based optimization, but lacked either differentiability or the ability to reason about belief-space uncertainty. Enter pdSTL (probabilistic differentiable Signal Temporal Logic), introduced by Bennett Dogbey and Hemanth Manjunatha in their latest arXiv preprint. pdSTL unifies probabilistic semantics with differentiable robustness over belief trajectories, using interval-valued probabilistic semantics to compute conservative satisfaction bounds. These bounds are propagated compositionally through the STL syntax tree, while the temporal robustness evaluation is formulated as a recurrent, LSTM-style unfolding of STL operators, enabling linear-time, differentiable monitoring suitable for end-to-end trajectory optimization.

The framework was validated on simulated obstacle avoidance and lane-change maneuvers, as well as real-world Crazyflie quadcopter flight experiments under aerodynamic disturbances. Results show pdSTL achieves efficient optimization with formal probabilistic guarantees, significantly outperforming deterministic differentiable STL in maintaining safety margins under real-world uncertainty. This work represents a step forward for safety-critical robotics, offering a scalable way to certify that autonomous systems will behave safely even when sensors and dynamics are imperfect. The paper is available on arXiv under subject areas Robotics and Systems and Control.