ScholarWorks@중앙대학교

초록

Service robots require instruction-following capabilities to perform various tasks, regardless of environmental changes. Thus, a task planner must accurately infer user intent even when human instructions are ambiguous. Furthermore, when initial plans become invalid due to unpredictable environmental states, the planner should modify its low-level policies and update target objects accordingly. To this end, we propose TIGER, a new task planning framework that generates highly reliable action sequences by deriving immutable subgoals from instructions. It also improves action quality by correctly selecting objects that match the user’s intent. Moreover, a task-representative one-shot strategy improves subgoal generation. Although TIGER uses ten times fewer examples for subgoal generation than a previous in-context learning approach using a large language model (LLM), it significantly improves high-level planning performance. Specifically, TIGER outperformed the previous LLM-Planner in the ALFRED benchmark, increasing success rates from 15.09 % to 35.06 % on the seen set and from 19.73 % to 42.57 % on the unseen set. TIGER also successfully executed ambiguous instructions in multi-room layouts with cluttered scenes, even without prior knowledge of the environment. Furthermore, its scalability was verified in real-world experiments.

키워드