Abstract: The present disclosure provides a Reinforcement Learning (RL) based architecture to efficiently learn action embeddings in low dimensional space. In conventional methods, the embeddings are learnt with the sole objective of improving policy learning, and there are no specific requirements on the quality of the embeddings. Initially, the system receives a goal to be reached by a mobile robot and a current location of the mobile robot is obtained. Simultaneously current transition dynamics associated with the plurality of directional actuators are obtained using a Reinforcement Learning (RL) technique. Further, a plurality of embeddings is computed based on the current location of the mobile robot and the current transition dynamics using a trained Dual Channel Training (DCT) based autoencoder decoder model. Finally, a displacement vector for current navigation of the mobile robot is computed based on the computed plurality of embeddings using the RL technique.