Abstract-to-Executable Trajectory Translation for One-Shot Task Generalization

• DOI: 10.48550/arxiv.2210.07658
• Publication Date: 2022-01-01
• Authors: Stone Tao; Xiaochen Li; Tongzhou Mu; Zhiao Huang; Yuzhe Qin; Hao Su
• Keywords: Executable

Training long-horizon robotic policies in complex physical environments is essential for many applications, such as manipulation. However, learning a policy that can generalize to unseen tasks challenging. In this work, we propose achieve one-shot task generalization by decoupling plan generation and execution. Specifically, our method solves three steps: build paired abstract environment simplifying geometry physics, generate trajectories, solve the original an abstract-to-executable trajectory translator. environment, dynamics manipulation are removed, making trajectories easier generate. introduces large domain gap between actual executed lack low-level details not aligned frame-to-frame with trajectory. manner reminiscent of language translation, approach leverages seq-to-seq model overcome executable enabling follow Experimental results on various different robot embodiments demonstrate practicability methods generalization.