Integrated Safe Actor-Critic Control with Adaptive Replay and Uncertainty
An arXiv preprint (2608.04732) introduces a unified architecture aimed at safe actor-critic control, which merges barrier filtering, uncertainty estimation, and experience replay into one cohesive system. This methodology adjusts the obstacle geometry within a control barrier function according to uncertainty assessments, employs filter interventions alongside estimation residuals to prioritize replay, and trains the critic based on actual actions taken instead of idealized ones. The approach was evaluated through a two-dimensional robot-navigation challenge featuring corrupted obstacle data, comparing six configurations with matched components across shared training budgets, random seeds, sensor inputs, exploration, and disturbances. Tests included a moderate post-training evaluation, an eleven-level perception-noise assessment, and an extreme-stress scenario at a 6.0 multiplier, where the integrated setup reported no contacts. This paper is classified as a cross-type announcement and is accessible on arXiv.
Key facts
- The paper is titled 'Toward Integrating Adaptive Experience Replay and Online Uncertainty Estimation in Safe Actor-Critic Optimal Control'.
- It is available on arXiv with identifier 2608.04732.
- The architecture integrates barrier filtering, uncertainty estimation, and experience replay.
- Uncertainty estimates update obstacle geometry for the control barrier function.
- Filter interventions and estimation residuals determine replay priority.
- The critic learns from executed rather than nominal actions.
- The method was tested on a 2D robot-navigation task with corrupted obstacle measurements.
- Six component-matched configurations were compared under common training budgets, random seeds, sensor streams, exploration, and disturbances.
- Evaluations included a moderate post-training test, an eleven-level perception-noise sweep, and an extreme-stress test at multiplier 6.0.
- In the extreme test, the integrated configuration recorded no contacts.
Entities
Institutions
- arXiv