Answers are grounded in this paper and cite their sources.
Demonstrates improved action prediction in offline reinforcement learning, implying better performance under corrupted observations and delays.
Offline reinforcement learning (offline RL) learns policies from pre-collected trajectories without online interaction with the environment, making it suitable for high-cost and safety-critical control applications. However, offline training datasets may suffer from observation corruption, while deployed agents may experience observation delay, resulting in distorted trajectory contexts and mismatches between available observations and current actions. To address this compound degradation, this paper proposes the Observation Recovery and Belief-Integrated Transformer (ORBIT), a robust offline RL algorithm that integrates diffusion-model-based observation recovery with delay-aware belief-based decision-making. ORBIT first employs an Observation Recovery Module (ORM) to detect and selectively recover suspicious observations by exploiting local temporal consistency. It then uses a Delay-Aware Belief Decision Transformer (DBDT) to infer decision-relevant information from delayed observations, historical actions, previously observed rewards, and delay-step information. The inferred belief representation is fused with delayed-observation features and fed into a Decision Transformer backbone for action prediction. Experiments on D4RL MuJoCo tasks show that ORBIT achieves an average normalized score of 32.1 in corrupted no-delay settings, outperforming Decision Transformer (DT) and Robust Decision Transformer (RDT) by 52.9% and 43.9%, respectively. Under adversarial corruption with fixed observation delays, ORBIT achieves 15.8, outperforming Delay-Transformer Belief Policy Constrained Offline Reinforcement Learning (DT-CORL) and RDT.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: