Multi-Task Policy Learning with Key Region Extraction for Visual Understanding

Le résumé fourni par la source

In complex multi-task scenarios, training robots often necessitates large amounts of expert data and involves processing high-dimensional multimodal observation spaces. This poses significant challenges to the development of general-purpose robotic systems, especially when deployed in real-world environments, where additional constraints and low-latency requirements must also be considered. Therefore, it is crucial to develop a framework that can efficiently leverage expert data while addressing the challenges associated with high-dimensional observation spaces. This paper proposes a transformer-based architecture that utilizes offline imitation learning to enable robotic arms to learn policies more efficiently. By incorporating image keypoint extraction and segmentation of critical regions within high-dimensional observations, the model facilitates faster and more efficient processing of multimodal data. Experiments were conducted in simulation using MetaWorld, covering 30 multitask scenarios, with an overall success rate of 82 %. Additionally, real-world experiments were carried out across eight distinct scenarios, with approximately 20 expert data collected per scenario. The success rates ranged from 50 % to 70 %, indicating notable improvements over other algorithms.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.