FISRF: Flexible Instructional Scaffolding Reflection Framework for Robotics Task Planning with Large Language Model

Qingmiao Liang, Jiatao Zhang, Runxin Zou, Yufan Song, Wei Song, Yuhua ZhengarXiv 2026 · arXiv

TL;DR

FISRF (Flexible Instructional Scaffolding Reflection Framework) is a novel agent-mentor architecture that brings educational psychology to robot task planning. By identifying when a robot is stuck in its "Zone of Proximal Development," the framework adaptively switches between autonomous self-reflection and targeted mentor guidance, reducing error-correction rounds by 37% and improving overall success rates.

What Problem Does This Paper Address?

Large Language Models (LLMs) have become powerful engines for robotic task planning, but they aren't perfect. When faced with new environments, these agents often fall into "hallucination loops" or logical inconsistencies.

Existing "self-reflection" methods—where the robot looks at its own failure and tries to fix it—rely almost entirely on autonomous exploration. In unfamiliar territory, this often leads to the accumulation of "worthless experience": the robot fails, reflects poorly, fails again in a similar way, and eventually runs out of permitted steps. The authors identify a critical gap: robots lack a mechanism to recognize when they are truly stuck and need a "nudge" from a more knowledgeable source to turn a failure into a learning moment.

Methods

The core of FISRF is the Agent-Mentor architecture, inspired by the concept of "Instructional Scaffolding." The framework consists of four main components:

  1. The Actor: An LLM that generates the actual action sequences.
  2. The Evaluator: The "brain" that decides if the robot needs help. It calculates two key metrics: Difficulty of Correction (DC) and Magnitude of Improvement (MI).
  3. Self-Reflection Module: Distills failed trajectories into reusable lessons.
  4. The Mentor: A Retrieval-Augmented Generation (RAG) module that provides state-conditioned guidance from a structured knowledge base of successful task examples.
The FISRF framework architecture showing the interaction between the Agent, Mentor, and Memory modules.

The "magic" happens when the Evaluator determines the robot has entered the Zone of Proximal Development (ZPD)—the gap between what it can do alone and what it can do with help. If the robot's improvement is low ($MI \leq 50\%$) or the correction is too difficult ($DC > 0.5$), the Mentor intervenes with specific instructions to get the robot back on track.

Datasets

The researchers evaluated FISRF using:

  • ALFWorld: A text-based virtual household simulator. They tested 134 tasks across six categories: Put, Clean, Heat, Cool, Examine, and Pick & Place.
  • Real-World Robotics: Validation on a dual-arm mobile robot platform performing kitchen item manipulation tasks to prove the framework's transferability to physical hardware.

Key Insights and Ideas

  • Moving Beyond Fixed Reflection: Traditional methods use a "one-size-fits-all" reflection prompt. FISRF proves that reflection should be dynamic; simple errors need "light intensity" reflection, while persistent failures require "high intensity" mentor intervention.
  • Quantifying "Stuckness": By mathematically modeling the ZPD using task completion rates and sub-goal progress, the authors provide a way for agents to "know what they don't know."
  • Scaffolding vs. Hand-holding: The mentor doesn't just give the answer; it provides "scaffolding" that the robot eventually internalizes. As the robot gains experience, the need for external support decreases, leading to more autonomous competence.
  • Efficiency is Key: The framework doesn't just aim for success; it aims for efficient success. By reducing redundant reasoning, it saves significant computational cost (tokens).
A comparison of FISRF against traditional fixed-strength reflection and complexity-based reflection.

Key Results

FISRF consistently outperformed state-of-the-art baselines like Reflexion and FCRF:

  • Success Rate: Achieved a 94.8% overall success rate (using GPT-4o), with a perfect 100% on "Put" and "Cool" tasks.
  • Efficiency: Reduced the average number of error-correction rounds by 37% and redundant reasoning by 25.1%.
  • Faster Convergence: As shown in the episode analysis, FISRF reaches high success rates much earlier than other methods, demonstrating its ability to learn rapidly from a single failure.
Success rate and revision round curves showing FISRF's superior convergence speed across episodes.

Limitations and Future Work

While FISRF is a significant step forward, the authors acknowledge a few areas for growth:

  • Static Knowledge Base: The Mentor currently relies on a pre-constructed, offline knowledge base. Future versions could involve an "automatic knowledge-base expansion" where the robot adds its own successful recoveries to the Mentor's library.
  • Cross-Platform Transfer: The current study focuses on specific environments; investigating how mentor guidance can be transferred across vastly different robotic platforms (e.g., from a kitchen bot to a warehouse bot) remains an open challenge.
  • Real-world Latency: While the framework reduces token counts, the multi-LLM architecture (Actor, Evaluator, Mentor) adds complexity that needs to be optimized for real-time robotic response.

Summary updated

Abstract

Large Language Models (LLMs) are widely adopted for reflective error correction in task planning; however, existing approaches primarily rely on autonomous exploration, rendering them prone to accumulating erroneous experiences during early interactions with new environments. To address this limitation, we propose the Flexible Instructional Scaffolding Reflection Framework (FISRF), an agent--mentor architecture that estimates the robot's Zone of Proximal Development (ZPD) in task planning and adaptively selects reflection strategies and mentor guidance, thereby facilitating the accumulation of effective error-correction experience. Experiments in ALFWorld and on real-world robotic platforms demonstrate that FISRF reduces the average number of error-correction rounds by 37% and redundant reasoning by 25.1%, while improving the overall task success rate by 8%. Project webpage : https://fisrfplanning.github.io/
BibTeX
@misc{liang2026fisrf,
  author = {Qingmiao Liang and Jiatao Zhang and Runxin Zou and Yufan Song and Wei Song and Yuhua Zheng},
  title = {{FISRF: Flexible Instructional Scaffolding Reflection Framework for Robotics Task Planning with Large Language Model}},
  year = {2026},
  eprint = {2610.35686},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35686},
}
Figure