DyRef: Dynamic Reflection Framework Via Graph-Based Complexity for Robotic Planning
Jiatao Zhang, QingMiao Liang, Tuocheng Hu, Yufan Song, Wei Song, Shiqiang Zhu
AI summary
Problem
Existing LLM-based robotic planning methods rely on fixed reflection routines that ignore multi-dimensional task complexity, causing redundant reflections or insufficient error correction.
Approach
The framework converts tasks and trajectories into a Diagnostic Graph to quantify complexity across four dimensions, then uses a learned routing policy to dynamically select and combine tailored reflection tools.
Key results
- Diagnostic Graph modeling dependency, spatial, interaction, and structural complexity
- Self-supervised routing policy dynamically selecting reflection tools
- 16.1% improvement in first-trial success rates on AlfWorld and real robots
- 64.4% reduction in redundant reflections
Why it matters
Provides a principled, adaptive reflection mechanism that significantly improves planning reliability and efficiency for LLM-driven robotic systems.
Abstract
Robotic planning tasks often involve diverse com- plexities, which make adaptive improvement through reflection particularly challenging. Existing LLM-based approaches typi- cally rely on fixed routines, lacking the ability to adjust to task- specific complexity and often leading to redundant reflections. To address this, we propose DyRef, a dynamic reflection frame- work that models tasks as a Diagnostic Graph, measures task complexity through structural factors, and routes them through a Reflection Toolkit via a learned Routing Policy network. This design enables tailored reflection strategies that reduce redundancy and improve reasoning efficiency. Experiments in AlfWorld and on real-world robotic platforms show that DyRef improves first trial success rates by 16.1%, while reducing redundant reflections by 64.4%.