[PDF] Multi-step planning with learned effects of (possibly partial) action executions-论文阅读讨论-ReadPaper

摘要

In this paper, we propose an affordance model, which is built on ConditionalNeural Processes, that can predict effect trajectories given objects, action oreffect information at any time. Affordances are represented in a latentrepresentation that combines object, action and effect channels. This modelallows us to make predictions of intermediate effects expected to be obtainedfrom partial action executions, and this capability is used to make multi-stepplans that include partial actions in order to achieve goals. We first showthat our model can make accurate continuous effect predictions. We compared ourmodel with a recent LSTM-based effect predictor using an existing dataset thatincludes lever-up actions. Next, we showed that our model can generate accurateeffect predictions for push and grasp actions. Finally, we showed that oursystem can generate successful multi-step plans in order to bring objects todesired positions. Importantly, the proposed system generated more accurate andeffective plans with partial action executions compared to plans that onlyconsider full action executions. Although continuous effect prediction andmulti-step planning based on learning affordances have been studied in theliterature, continuous affordance and effect predictions have not been utilizedin making accurate and fine-grained plans.

展开全部

图表提取

论文十问

参考文献

被引用

提问与回答

暂无人提供速读十问回答