Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Agentic Self-Evolutionary Replanning for Embodied Navigation

Guoliang Li, Ruihua Han, Chengyang Li, He Li, Shuai Wang, Wenchao Ding, Hong Zhang, Chengzhong Xu

TL;DR

This work proposes agentic self-evolving action model that uses in-context learning with auto-differentiation (ILAD) for adaptive function adjustment and global parameter reset and introduces agentic self-evolving action model that uses in-context learning with auto-differentiation (ILAD) for adaptive function adjustment and global parameter reset.

Abstract

Failure is inevitable for embodied navigation in complex environments. To enhance the resilience, replanning (RP) is a viable option, where the robot is allowed to fail, but is capable of adjusting plan until success. However, existing RP approaches freeze the ego action model and miss the opportunities to explore better plans by upgrading the robot itself. To address this limitation, we propose Self-Evolutionary RePlanning, or SERP for short, which leads to a paradigm shift from frozen models towards evolving models by run-time learning from recent experiences. In contrast to existing model evolution approaches that often get stuck at predefined static parameters, we introduce agentic self-evolving action model that uses in-context learning with auto-differentiation (ILAD) for adaptive function adjustment and global parameter reset. To achieve token-efficient replanning for SERP, we also propose graph chain-of-thought (GCOT) replanning with large language model (LLM) inference over distilled graphs. Extensive simulation and real-world experiments demonstrate that SERP achieves higher success rate with lower token expenditure over various benchmarks, validating its superior robustness and efficiency across diverse environments.

Agentic Self-Evolutionary Replanning for Embodied Navigation

TL;DR

This work proposes agentic self-evolving action model that uses in-context learning with auto-differentiation (ILAD) for adaptive function adjustment and global parameter reset and introduces agentic self-evolving action model that uses in-context learning with auto-differentiation (ILAD) for adaptive function adjustment and global parameter reset.

Abstract

Failure is inevitable for embodied navigation in complex environments. To enhance the resilience, replanning (RP) is a viable option, where the robot is allowed to fail, but is capable of adjusting plan until success. However, existing RP approaches freeze the ego action model and miss the opportunities to explore better plans by upgrading the robot itself. To address this limitation, we propose Self-Evolutionary RePlanning, or SERP for short, which leads to a paradigm shift from frozen models towards evolving models by run-time learning from recent experiences. In contrast to existing model evolution approaches that often get stuck at predefined static parameters, we introduce agentic self-evolving action model that uses in-context learning with auto-differentiation (ILAD) for adaptive function adjustment and global parameter reset. To achieve token-efficient replanning for SERP, we also propose graph chain-of-thought (GCOT) replanning with large language model (LLM) inference over distilled graphs. Extensive simulation and real-world experiments demonstrate that SERP achieves higher success rate with lower token expenditure over various benchmarks, validating its superior robustness and efficiency across diverse environments.
Paper Structure (13 sections, 8 equations, 10 figures, 4 tables, 1 algorithm)

This paper contains 13 sections, 8 equations, 10 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. System Overview
  4. Problem Formulation
  5. Cross-level Replanning
  6. Agentic Self-Evolutionary Replanning
  7. Local Agentic Self-Evolution
  8. Global Graph Chain-of-Thought
  9. Experiments and Results
  10. Qualitative Evaluations of SERP
  11. Quantitative Evaluations of SERP
  12. Real-World Experiment
  13. Conclusion

Figures (10)

  • Figure 1: SERP replans at two-levels with self-evolution.
  • Figure 2: Overview of SERP, which consists of local agentic self-evolution and global graph chain-of-thought.
  • Figure 3: Local agentic self-evolution.
  • Figure 4: Parameter retrieval in 00824.
  • Figure 5: Evaluation of Local replanning: (a) AD with proper RAG; (b) AD with poor RAG; (c) SERP with poor RAG; (d) parameter heatmap of (c).
  • ...and 5 more figures