Integrated safe actor-critic control hits zero collisions in extreme robot tests

A new arXiv paper (2608.04732) targets a persistent flaw in safe actor-critic control: barrier filtering, uncertainty estimation, and experience replay are typically designed as independent modules, even though all three reshape the data used for learning. Authors Mahshad Rastegarmoghaddam, Davoud Nikkhouy, and Shima Samadzadeh propose an integrated architecture where the online uncertainty estimate continuously updates the obstacle geometry used by a control barrier function, filter interventions and estimation residuals drive replay priority, and the critic trains on executed rather than nominal actions. This coupling lets the system adapt its safety constraints based on real perception confidence, rather than relying on static barriers that fail under sensor corruption.

The team validated the approach on a 2D robot-navigation task with corrupted obstacle measurements, comparing six component-matched configurations under identical training budgets, seeds, and sensor noise. Performance was evaluated in a moderate post-training test, an eleven-level perception-noise sweep, and an extreme stress test at a 6.0x noise multiplier. In the extreme test, the integrated configuration achieved zero contacts and reached the goal across all five evaluation seeds, posting a mean cost of 7.63±0.44 and an obstacle-belief RMSE of 3.52±0.55 cm. An ablation with uncertainty estimation but no full integration also achieved zero contacts but reached the goal in only four of five seeds, with mean cost 8.96±2.08 and belief error 11.08±1.23 cm—about three times worse. The paper also derives a finite-training bound for replay exposure and a robust barrier condition specifying required estimation-error and feasibility assumptions. The authors are careful to note that broader safety and convergence guarantees require further study, but the benchmark strongly supports coupling estimation, safety filtering, and replay in safety-critical control.