Safe Continual Domain Adaptation after Sim2Real Transfer of Reinforcement Learning Policies in Robotics

TL;DR

The paper addresses the reality gap in robotic reinforcement learning by enabling safe, continual adaptation after deployment through Safe Continual Domain Adaptation (SCDA). SCDA combines domain randomization with a safe RL framework (PCRPO) and Elastic Weight Consolidation (EWC) to first learn a robust policy in a highly randomized simulator and then adapt safely to the real system while remembering the pretraining policy. The approach regularizes adaptation via the diagonal Fisher information to prevent forgetting important parameters and enforces safety through constrained RL during pretraining and a safety-aware objective during real-world updates, yielding better performance across targets without catastrophic forgetting. Empirically, SCDA improves over zero-shot transfer and other baselines on reach/balance and grasping tasks, demonstrating safer and more reliable post-deployment adaptation with practical implications for autonomous robotic systems.

Abstract

Domain randomization has emerged as a fundamental technique in reinforcement learning (RL) to facilitate the transfer of policies from simulation to real-world robotic applications. Many existing domain randomization approaches have been proposed to improve robustness and sim2real transfer. These approaches rely on wide randomization ranges to compensate for the unknown actual system parameters, leading to robust but inefficient real-world policies. In addition, the policies pretrained in the domain-randomized simulation are fixed after deployment due to the inherent instability of the optimization processes based on RL and the necessity of sampling exploitative but potentially unsafe actions on the real system. This limits the adaptability of the deployed policy to the inevitably changing system parameters or environment dynamics over time. We leverage safe RL and continual learning under domain-randomized simulation to address these limitations and enable safe deployment-time policy adaptation in real-world robot control. The experiments show that our method enables the policy to adapt and fit to the current domain distribution and environment dynamics of the real system while minimizing safety risks and avoiding issues like catastrophic forgetting of the general policy found in randomized simulation during the pretraining phase. Videos and supplementary material are available at https://safe-cda.github.io/.

Figures (5)

• Figure 1: A conceptual representation of Safe Continual Domain Adaptation (SCDA). The unknown real system distribution (blue X) is approximated in simulation by randomizing the system parameters and training a robust RL policy (dashed square) covering wide parameter ranges, while simultaneously learning a safety Critic (shaded area). Then, the policy is adapted to the real system, using continual learning to prevent diverging from the initial robust policy and the safety Critic to avoid unsafe, exploitative actions in the real system. If real-world system parameters change (pink X) the policy can adapt to accommodate the change.
• Figure 2: Reaching task -- simulation and real system environment.
• Figure 3: Training progress during the pretraining phase (top) and the adaptation phase on the realistic (middle) and real system (bottom) under different adaptation strategies in terms of the average timestep reward (left), average timestep cost (center), and the sum of both (right) while adapting to the current target. The target switches every ten iterations during the adaptation phase. Results are averaged over five seeds.
• Figure 4: Realistic (top) and real system (bottom): Performance of different adaptation strategies in terms of the episode reward (left), cost (center), and total (right) while adapting to the current target (Current) compared to performance on the other targets (Others). Results are shown for five seeds.
• Figure 5: Relative parameter change as a function of Fisher importance.