Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: Today's paper: "Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning".
Dev: The gist ARMOR introduces an adaptive round-based multi-task model for robotic failure detection and reasoning that iteratively refines its predictions by jointly predicting detection outcomes and natural language reasoning,
Rosa: First, who's behind it and why it matters.
Paper summary: Dev: So looking at Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning, the title itself really sums up what they are doing with this ARMOR model.
Rosa: They moved away from treating failure reasoning as just a closed-set classification problem, which is what most prior work did. Instead, they focus on detection as one thing and reasoning as an open-ended language modeling problem that refines itself round by round.
Taro: What this means for the field is that we might be able to detect subtle failures in complex robotic tasks without needing thousands of perfectly labeled examples for every single failure type.
Dev: It suggests that we don't need to pre-define every possible thing a robot can do wrong before we build a robust system. We just need the iterative refinement process to figure it out as it goes.
Rosa: The authors show that this method actually improves performance by up to thirty percent on failure detection rate and up to one hundred percent in reasoning measured through an LLM fuzzy match score compared to previous approaches.
Taro: That improvement in reasoning quality, especially when it's open-ended, is significant because it shows the model can capture those real-world failures beyond the fixed modes they mentioned.
Dev: The authors also point out that this method handles data heterogeneity better than others by learning from both sparse binary labels and small sets of rich annotations.
Rosa: They acknowledge a limitation in their work—they noted that reasoning results tend to be corrected based on detection, since detection is generally more reliable because there's more data for it.
Taro: And they also pointed out that the method relies on having good initial training signals to start that refinement process effectively.
Dev: So ultimately, Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning is about moving from rigid classification to a dynamic, self-improving system for understanding robot failures.
Conclusion: Rosa: So, we're wrapping up this look at ARMOR, which stands for Adaptive Round-based Multi-task Model for Robotic failure detection and reasoning.
Dev: Yeah, it’s about taking those models that can see a robot and talk about what’s wrong with it—the detection part—and making them actually reason about *why* things went wrong.
Taro: The core idea is that instead of just guessing a failure category, the model keeps predicting the failure and also generating a natural language explanation for it, over and over again.
Rosa: Iteratively, right? So if it guesses a problem in round one, it uses that prediction to guide its guess in round two.
Dev: Exactly. It’s this self-refinement process where the model refines its own detection and its own reasoning language based on what it just said.
Taro: And they trained this thing using mixed data—those huge sets of simple failure labels and those smaller sets of detailed human reasoning examples.
Rosa: That’s the trick, right? It lets the system learn from both quick, sparse data and the expensive, rich annotations at the same time.
Dev: The results show that this iterative approach actually helps detection rates go up by about thirty percent and reasoning scores by a hundred percent in some tests.
Taro: That jump in reasoning score is interesting because it suggests it can handle failures that aren't pre-programmed, just based on the context.
Rosa: It means for real robotics, we might not need to manually list every possible way a robot could break; the system figures out the failure path.
Dev: But we gotta remember they flagged that this correction usually happens because it trusts its detection head more than its reasoning head initially.
Taro: That’s a good caveat—the detection part gets more reliable as you feed it more data, which then helps refine the language part.
Rosa: So, while it’s powerful for open-ended failures right now, we still need to see how stable this refinement loop is when the real world throws something completely unexpected at it.
Dev: That’s exactly where we need to look next—what happens when the system starts failing in a way that confuses both detection and reasoning simultaneously.
University of Texas at Austin · Amazon Robotics Institute of Technology and Research Center of Carnegie Mellon University
cs.RO, cs.LG
Submitted: 2026-02-12
Updated: 2026-10-15
Code: https://github.com/2U1/Qwen2-VL-Finetune
Project page: https://sites.google.com/utexas.edu/armor
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 89/100
The gist: The gist ARMOR introduces an adaptive round-based multi-task model for robotic failure detection and reasoning that iteratively refines its predictions by jointly predicting detection outcomes and
Key concepts
- Multi-task Self-Refinement Process
- This is the core idea where the model doesn't just make one prediction; it repeatedly predicts a detection outcome (failure or success) and then uses that result to generate a better natural language explanation. It loops back on itself, improving both its detection accuracy and the quality of its reasoning in each cycle.
- Heterogeneous Supervision
- The model is trained using two different types of data simultaneously: large amounts of simple binary labels (like 'failure' or 'success') which are easy to get automatically from sensors, and smaller amounts of high-quality, expensive human annotations for detailed reasoning. This allows the model to learn robustly from both simple and complex supervision.
- Adaptive Round-based Training
- Training occurs in rounds where the model refines its understanding. It starts with initial predictions and then uses those predictions to guide subsequent refinement steps. This iterative approach ensures that the reasoning becomes increasingly accurate over successive iterations, moving from a rough guess to a precise explanation.
Terminology
Summary
The gist
ARMOR introduces an adaptive round-based multi-task model for robotic failure detection and reasoning that iteratively refines its predictions by jointly predicting detection outcomes and natural language reasoning, enabling open-ended reasoning beyond fixed categories.
How it works
-
ARMOR formulates detection and reasoning as a
multi-task self-refinement process, where the model iteratively predicts detection outcomes and natural language reasoning conditioned on past outputs
(Page 2). -
The model is trained using heterogeneous supervision, combining
large-scale sparse binary labels and small-scale rich reasoning annotations - optimized via a combination of offline and online imitation learning
(Page 4). -
The architecture employs a
multi-task architecture with separate heads for detection and reasoning,
where a classification head predicts detection and the original language model head is retained for reasoning generation (Page 4). -
During training, the process involves two phases:
Phase I: Offline Imitation
where warm-up and expert-conditioned training occur, followed byPhase II: Online Refinement
usingonline rollouts from the policy
(Page 4). -
ARMOR's inference procedure is an iterative self-refinement that generates multiple refinement trajectories and selects the most confident prediction via a metric based on
self-certainty
(Page 6).
Key Contributions and Findings
-
The work proposes
ARMOR, which stands for adaptive round-based multi-task refinement of visionlanguage models (VLMs) for robotic failure detection and reasoning
(Page 4). -
ARMOR is the first to extend iterative reasoning with a VLM on multi-modal failure detection in robotics using video feed (Page 4).
-
Experiments show that ARMOR achieves performance improvements by
up to 30% on failure detection rate and up to 100% in reasoning measured through LLM fuzzy match score
compared to previous approaches (Page 1). -
The method is robust to data heterogeneity, as it handles the setting where
binary success/failure labels can often be obtained automatically (e.g., from robot sensors or system logs), whereas natural-language reasoning labels require expert annotation and are far more costly
(Page 3). -
Ablation studies confirm that combining all components in ARMOR yields the best performance, demonstrating the necessity of
multi-task design, refinement, and offline and online training
(Page 8).
Performance Metrics and Results
-
Performance is evaluated using three complementary metrics:
Binary Success Rate for detection,
LLM Fuzzy Match (semantic similarity),
andROUGE-L score
for reasoning quality (Page 19). -
ARMOR achieves state-of-the-art results across diverse environments, outperforming baselines like Qwen2.5-VL and Claude-3.7 on benchmarks such as RLBench, Sparrow, and Maniskill (Page 19).
-
Refinement significantly improves reasoning quality; for instance, in one setting,
Round 0 0.475±0.016
increased toRound 1 0.676±0.025 (+42.4%)
(Page 9). -
The iterative process demonstrates convergence:
In later rounds, it first adjusts the reasoning to stay consistent with the detection result (failure) and then refines it into an accurate account of the failure
(Page 10).
Limitations and Future Work
-
A limitation noted is that naive single-task generation limits accuracy; ARMOR addresses this by treating tasks as a
round-based multitask refinement problem
(Page 2). -
The paper acknowledges that "reasoning results tend to be corrected based on detection, since detection is generally more reliable (as more data is available for detection), though there are also cases where both predictions are corrected through successful refinement" (Page 10).
-
Future work includes incorporating
richer training signals such as task rewards and human preferences to further enhance reasoning quality and alignment with human
(Page 10). -
Extending ARMOR to other synergistic tasks, such as
predicting safety violations and recovery strategies, could enhance its applicability in real-world robotics
(Page 10).
How it works
-
ARMOR formulates detection and reasoning as a
multi-task self-refinement process, where the model iteratively predicts detection outcomes and natural language reasoning conditioned on past outputs
(Page 2). -
The model is trained using heterogeneous supervision, combining
large-scale sparse binary labels and small-scale rich reasoning annotations - optimized via a combination of offline and online imitation learning
(Page 4). -
The architecture employs a
multi-task architecture with separate heads for detection and reasoning,
where a classification head predicts detection and the original language model head is retained for reasoning generation (Page 4). -
During training, the process involves two phases:
Phase I: Offline Imitation
where warm-up and expert-conditioned training occur, followed byPhase II: Online Refinement
usingonline rollouts from the policy
(Page 4).
Improvements for AI systems
- Bold header: Multi-task Architectural Design for Joint Prediction
The system can jointly predict detection l via a classification head and reasoning e via a language decoder,
which addresses the challenge of combining classification-style detection with open-ended reasoning
by ensuring that both tasks are optimized appropriately while allowing their outputs to inform one another.
- Bold header: Iterative Self-Refinement for Uncertainty Reduction
ARMOR can iteratively refine its predictions
by generating multiple refinement trajectories and selecting the most confident prediction using a metric based on self-certainty, which is calculated as C(m) = H(m)det + λ H(m)reason.
- Bold header: Robust Training with Heterogeneous Supervision
The model can be trained effectively under data imbalance by leveraging offline imitation learning
from large-scale sparse labels and optimizing this via online refinement,
which allows the model to learn from both binary and rich reasoning annotations simultaneously.
- Bold header: Open-Ended Failure Reasoning
Unlike prior work that treats failure reasoning as a closed set of modes, ARMOR can perform open-ended, iterative refinement by jointly refining its detection and reasoning predictions,
enabling human-like reasoning that capture real-world failures beyond fixed categories.
- Bold header: Scalability with Model Capacity
The system's performance scales effectively with larger models, as shown by the observation that ARMOR achieves consistent performance gains compared to the 7B version
when using the 32B version of Qwen2.5-VL,
demonstrating that the approach is robust to increased model capacity.
- Bold header: Improved Reasoning Quality via Refinement Rounds
The system can achieve significant reasoning gains through iterative refinement, as demonstrated by Table 3 showing that Round 1
results in a substantial improvement in reasoning quality (e.g., +42.4%
), confirming that the iterative process reduces uncertainty and enhances both detection and reasoning quality.
Sources
- Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning
- Qwen2.5-VL Technical Report
- Constitutional AI: Harmlessness from AI Feedback
- Teaching Large Language Models to Self-Debug
- Vision-Language Models as Success Detectors
- GPT-4o System Card
- Scalable Best-of-N Selection for Large Language Models via Self-Certainty
- Learning Robust Reasoning through Guided Adversarial Self-Play
- REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction
- Show Your Work: Scratchpads for Intermediate Computation with Language Models
- Hierarchical Text-Conditional Image Generation with CLIP Latents
- Direct Nash Optimization: Teaching Language Models to Self-Improve with General Preferences
- RePLan: Robotic Replanning with Perception and Language Models
- LLaMA: Open and Efficient Foundation Language Models
- Voyager: An Open-Ended Embodied Agent with Large Language Models
- InternVideo: General Video Foundation Models via Generative and Discriminative Learning
- Qwen3 Technical Report
Related papers
- FMT x: An Efficient and Asymptotically Optimal Extension of the Fast Marching Tree for Dynamic Replanning
- MPCFormer: A physics-informed data-driven approach for explainable socially-aware autonomous driving
- RoboLab: A High-Fidelity Simulation Benchmark for Analysis of Task Generalist Policies
- ARC: A Reasoning Recipe for Robot Foundation Models
- HRDexDB: A 4D Dexterous Grasping Dataset Across Human and Multiple Robot Embodiments
- APT: Action Expert Pretraining Improves Instruction Generalization of Vision-Language-Action Policies