Adversarial Fine-tuning in Offline-to-Online Reinforcement Learning for Robust Robot Control

Researchers developed an offline-to-online reinforcement learning framework that improves robot control robustness through adversarial fine-tuning. The method trains policies on clean datasets then applies action perturbations during fine-tuning to build resilience against actuator faults and environmental uncertainties.

• →New framework combines offline efficiency with online adaptability for more robust robot control systems.
• →Adversarial fine-tuning with action perturbations significantly improves policy resilience against actuator faults.
• →Performance-aware curriculum balances robustness gains with nominal performance stability during training.
• →Method converges faster than training from scratch while outperforming offline-only approaches.
• →Results bridge the gap between sample-efficient offline learning and real-world deployment requirements.