Robust Robotic Task Planning via Immutable Subgoals
Chulyong Lim, Jaewon Baek, Junhee Han, WooYeol Bae, Woochul Nam
AI summary
Problem
LLM-based robotic planners frequently fail due to dependency violations, object hallucinations, and syntactic errors when handling ambiguous instructions in dynamic environments.
Approach
TIGER breaks instructions into environment-independent semantic subgoals and grounds them to real objects via a multi-stage visual pipeline, requiring only seven annotated examples.
Key results
- 35.06% seen and 42.57% unseen success rates on the ALFRED benchmark
- 14x reduction in annotation effort using only seven examples
- 76.7% to 100% success rate in real-world UR5e robot experiments
- Multi-stage visual grounding pipeline effectively localizes objects in cluttered scenes
Why it matters
Enables service robots to reliably execute complex instructions in unseen environments with minimal training data.
Abstract
Service robots require instruction-following capabilities to perform various tasks regardless of environmental changes. A task planner must accurately infer user intent even when human instructions are ambiguous. To this end, we propose TIGER, a task planning framework that generates reliable action sequences by deriving immutable subgoals from instructions. TIGER employs an Immutable Subgoal Planner (ISP) to decompose instructions into environment-independent subgoals and a Target Grounder (TG) to ground abstract