ADP: Adaptive Diffusion Policy Energizes Robots Thinking in Both Learning and Practice

Abstract

Adaptive control policies for robots often require balancing generalization from large offline datasets with efficient adaptation to specific deployment conditions. In this paper, we propose Adaptive Diffusion Policy (ADP), a two-stage framework that integrates temporal-aware diffusion models with parameter-efficient LoRA adaptation. First, in the learning stage, ADP imitates and generates actions based on image and video signals from a meager amount of expert demonstrations, considering both spatial and temporal information. This component contrasts with most existing works, which focus solely on spatial information. Second, in the practice stage, ADP incorporates a low-rank adaptation module into the policy, subsequently training it using residual reinforcement learning with minimal environment interaction. Experiments conducted on Meta-World benchmark demonstrate the efficiency of each ADP component and the superiority of ADP over representative baseline methods. Note to Practitioners—This work introduces Adaptive Diffusion Policy (ADP), a two-stage visuomotor framework that first learns from just a few image-and-video demonstrations by modeling both spatial and temporal cues, then rapidly refines its behavior via a lightweight low-rank adapter and residual reinforcement learning. The ADP enables swift skill acquisition on new tasks with minimal expert data and limited environment trials, making it ideal for industrial or household robots where extensive data collection is impractical. To apply ADP, collect a small set of demonstration clips, train the diffusion-based policy offline, and deploy the adapter online for in situ fine-tuning. The proposed Meta-World results show the ADP’s consistent gains over standard imitation and residual RL baselines, which is very easy for practitioners in multiple real-world robot scenarios.