Robust Memory Augmentation by Constrained Latent Imagination
The latent dynamics model summarizes an agent’s high dimensional experiences in a compact way. While learning from imagined trajectories by the latent model is confirmed to has great potential to facilitate behavior learning, the lack of memory diversity limits generalization capability. Inspired by a neuroscience experiment of “forming artificial memories during sleep”, we propose a robust memory augmentation method with Constrained Latent ImaginatiON (CLION) under a novel actor-critic framework, which aims to speed up the learning of the optimal policy with virtual episodic. Various experiments on high-dimensional visual control tasks with arbitrary image uncertainty demonstrate that CLION outperforms existing approaches in terms of data-efficiency, robustness to uncertainty, and final performance.
Code (0)
등록된 구현이 없습니다.
Tasks
DiversitySimilar Papers 제목 키워드 기반
Model-Based Reinforcement Learning via Imagination with Derived Memory
Model-based reinforcement learning aims to improve the sample efficiency of policy learning by modeling the dynamics of the environment. Recently, the latent dynamics model is further developed to enable fast planning in…
Model-based Reinforcement Learningreinforcement-learningReinforcement LearningReinforcement Learning (RL)MemoryVLA++: Temporal Modeling via Memory and Imagination in Vision-Language-Action Models
Temporal modeling is essential for robotic manipulation, as effective control requires both memory of past interactions and imagination of future states. However, most VLA models rely primarily on the current observation…
Koopman Dreamer: Spectrally Constrained Latent Dynamics for Stable World-Model Imagination
Latent world models improve sample efficiency in continuous control by optimizing policies over imagined latent trajectories, but common neural transitions offer limited direct control over modal persistence and error ac…
Continuous ControlFacing Off World Model Backbones: RNNs, Transformers, and S4
World models are a fundamental component in model-based reinforcement learning (MBRL). To perform temporally extended and consistent simulations of the future in partially observable environments, world models need to po…
Model-based Reinforcement LearningLatent Imagination Facilitates Zero-Shot Transfer in Autonomous Racing
World models learn behaviors in a latent imagination space to enhance the sample-efficiency of deep reinforcement learning (RL) algorithms. While learning world models for high-dimensional observations (e.g., pixel input…
Autonomous Racingcontinuous-controlContinuous ControlDeep Reinforcement Learning+1