Source-linked AI summary
REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction
Zeyi Liu, Arpit Bahety, Shuran Song
TL;DR
Robots need to detect and explain failures from complex, multimodal, long-horizon executions, but converting sensory experience into language reasoning remains challenging. REFLECT builds hierarchical summaries of robot experiences and progressively queries an LLM, then uses its explanations to guide correction planning. The framework generates informative explanations and achieves strong failure explanation, localization, and correction performance across simulation and real-world scenarios.
Problem
Generating textual summaries of multisensory robot data and systematically querying LLMs for failure reasoning remains challenging, limiting automated explanation of failed executions.
Method
REFLECT converts multisensory robot observations into a three-level hierarchical summary, progressively queries an LLM for failure reasoning, and guides a language-based planner to correct failures.
Results
REFLECT achieves the highest performance for explaining, localizing, and correcting failures in simulation and the real world, with around 80% correction planning success in simulation.
Takeaways & Limitations
Hierarchical multisensory summaries can support informative robot failure explanations and correction planning across several failure scenarios.
Takeaways & Limitations
The method assumes static environments and is less effective for low-level control failures because its summaries contain limited low-level state information.
Abstract
from arXiv · showhide
The ability to detect and analyze failed executions automatically is crucial for an explainable and robust robotic system. Recently, Large Language Models (LLMs) have demonstrated strong reasoning abilities on textual inputs. To leverage the power of LLMs for robot failure explanation, we introduce REFLECT, a framework which queries LLM for failure reasoning based on a hierarchical summary of robot past experiences generated from multisensory observations. The failure explanation can further guide a language-based planner to correct the failure and complete the task. To systematically evaluate the framework, we create the RoboFail dataset with a variety of tasks and failure scenarios. We demonstrate that the LLM-based framework is able to generate informative failure explanations that assist successful correction planning.
1 Introduction
REFLECT addresses the challenge of turning multisensory robot observations into language-based failure reasoning by organizing experiences hierarchically. It generates explanations that support human debugging and correction planning.
- Long-horizon robot failures require systems that can reflect on past experiences and explain failures in natural language.Such explanations can help users debug robots without reading execution logs and guide autonomous correction.
- Effective failure reasoning requires summarizing what happened, inferring what went wrong, and planning what to do next.These components respectively structure multimodal observations, assess expected plan outcomes, and support corrective task completion.
- The central challenge is generating textual summaries of robot sensory data and systematically querying LLMs for failure reasoning.The paper frames robot failure reasoning as a language reasoning task.
- Robot summaries should be multisensory because different failures may be easier to identify from visual, audio, or contact information.For example, auditory cues can reveal dropped objects or running water when visual evidence is less informative.
- Robot summaries should be hierarchical, combining high-level plan-outcome alignment with lower-level environmental context for explanation and correction.Higher levels localize failures, while lower levels preserve details needed for informative reasoning.
- REFLECT produces a three-level hierarchical summary and uses progressive failure explanation to guide a language-based planner toward correction plans.Experiments assess explanation quality with human evaluators and correction planning across several failure scenarios.
- RoboFail contains 100 simulated failure demonstrations and 30 real-world demonstrations collected with a UR5e robot arm.The dataset supports systematic evaluation across simulated and physical robot failures.
2 Related Work
Prior work studies robot failure explanation, video captioning, and LLM-based planning, but is constrained by narrow scenarios, limited robot-video data, or simplified feedback. REFLECT targets broader failure reasoning from multisensory, long-horizon experience summaries.
- Robot video captioning remains difficult because of limited data, although foundation models enable zero-shot captioning for some egocentric human activity videos.
- Robot failure explanation has been studied to increase human trust and help non-experts assist robots during failures.
- Earlier failure-explanation methods address specific scenarios such as picking, pick-and-place, navigation, or short-horizon manipulation.The cited studies therefore cover limited task and failure configurations.
- REFLECT uses LLM reasoning to detect and explain a wider range of failure scenarios without assumptions about task configuration or failure type.
- Prior LLM task-planning systems often rely on ground-truth feedback tied to one or a few actions, whereas REFLECT studies multisensory observations over long-horizon robot tasks.
3 Method: the REFLECT Framework
REFLECT converts multisensory robot experience into three abstraction levels, then progressively queries an LLM to localize and explain failures before generating correction plans. Its summaries connect task outcomes with detailed event evidence.
- Framework overview: REFLECT comprises hierarchical summarization, progressive failure explanation, and failure correction planning modules.
- Hierarchical summary: The hierarchical summary aggregates sensory data over time into sensory-input, event-based, and subgoal-based representations.
- Sensory-input summary: The sensory-input summary converts multimodal observations into unified information for failure explanation, including task-informed scene graphs and audio summaries.
- Sensory-input summary: Scene graphs represent task-relevant objects, their states, spatial relations, and the robot as a node with gripper relations.The graph is constructed from detected objects, depth-projected semantic point clouds, state labels, and spatial heuristics.
- Sensory-input summary: Audio processing identifies candidate events from segmented sound clips using audio-text embedding similarity.The selected label is the candidate with the highest cosine similarity.
- Event-based summary: The event-based summary selects key frames when scene graphs change, audio events begin or end, or a subgoal ends, then captions those frames.
- Subgoal-based summary: The subgoal-based summary stores end-of-subgoal observations so the LLM can compare execution outcomes with high-level plan expectations.
- Progressive failure explanation: Progressive explanation first verifies each subgoal, then analyzes event history for execution failures or the final state and original plan for planning failures.Execution analysis can identify an immediate cause such as dropping a pot, while planning analysis can expose mismatches between actions and goals.
4 The RoboFail Dataset
RoboFail is a mixed simulation and real-world dataset of robot failure demonstrations, with failures organized through a dedicated taxonomy. It supports evaluation across varied tasks and settings.
- Dataset construction: The simulated RoboFail portion contains 100 failure scenarios across 10 tasks, with 10 cases per task.It stores RGB-D, sound, robot-state observations, and simulation ground-truth metadata.
- Dataset construction: The real-world portion contains 30 failure scenarios across 11 tasks collected through human teleoperation of a UR5e arm in a toy kitchen.RGB-D observations and recorded sound are stored for these demonstrations.
- Failure taxonomy: RoboFail includes a taxonomy visualizing its failure scenarios.
5 Evaluation
REFLECT is evaluated on simulated and real-world robot failures using explanation, localization, and correction-planning measures, alongside ablations and qualitative comparisons. Results show benefits from progressive reasoning, hierarchical multisensory summaries, audio, and failure explanations for correction planning.
- Overall results: REFLECT achieves the highest performance in explaining, localizing, and correcting failures in both simulation and the real world.Performance decreases slightly in the real world due to perception errors.
- Ablations and analysis: Removing audio decreases explanation and localization accuracy by around 20% on execution failures compared with REFLECT.Audio helps detect unexpected events and visually occluded object states.
- Ablations and analysis: Progressive failure explanation improves localization and explanation by identifying the failed subgoal before querying event-based evidence for the cause.In the egg example, REFLECT identifies a failed “pick up egg” action and infers that the fridge was closed.
- Ablations and analysis: Event-based summaries improve failure explanation by preserving intermediate environmental observations, including auditory and visual evidence of a dropped pot.These observations help identify both the failure and its exact time step.
- Correction planning: Explanation-guided correction planning addresses failure causes, whereas planning without explanation can repeat the original plan without corrective actions.For a mug-placement failure, REFLECT moves an obstructing cup before proceeding.
6 Conclusion
REFLECT converts multisensory robot observations into hierarchical experience summaries, uses progressive LLM failure explanations, and evaluates them on failed executions in simulation and the real world.
- REFLECT converts multisensory observations into a hierarchical summary of robot past experiences and progressively queries an LLM for failure explanations.
- The generated explanations guide a language planner to correct failures and complete tasks.
- The RoboFail dataset contains failed robot executions from both simulation and the real world for framework evaluation.
- REFLECT achieves better performance than several baselines and ablations.
A Method Details
The method builds summaries from spatially and temporally aggregated observations, verifies subgoals, and generates correction plans using available robot actions.
- Visual and temporal summarization: Eight spatial-relation heuristics are computed from aggregated semantic point clouds and 3D object bounding boxes.
- Visual and temporal summarization: Objects are added, updated, replaced, or deleted in the accumulated point cloud according to observations, alignment, motion, and interaction.
- Failure analysis and correction: Subgoal verification prompts compare each goal with post-execution observations and enter execution analysis when a subgoal is unsatisfied.
- Failure analysis and correction: Correction planning uses available actions while constraining the robot to hold at most one object in its gripper.
A.3.2 Failure explanation: execution analysis
Execution analysis uses action and observation history to explain failures at identified time steps, including a dropped object that caused a later placement failure.
- The execution-analysis prompt requests a brief explanation of what happened and what caused the failure, with relevant time steps when possible.
- At 00:44, the robot attempted to place the pot on the fourth stove burner while the pot was no longer in its gripper.
- The failure was caused by the robot dropping the water-filled pot at 00:36 while moving toward the fourth stove burner.
- Failure time steps can be extracted from the generated failure explanation through a follow-up LLM prompt.
A.3.3 Failure explanation: planning analysis
Planning analysis diagnoses failures in otherwise successfully executed plans by comparing the intended goal with the final state and plan actions.
- When all subgoals are satisfied, planning analysis considers whether the original robot plan contains a mistake.
- The analysis uses the task description, goal condition, robot plan, and final state to identify what went wrong.
- The robot placed the pot on the fourth burner but turned on the second burner, creating a mismatch between the pot location and active burner.
- The most relevant failure time step for this planning failure is 00:49.
B.1 Dataset Details
RoboFail contains simulation and real-world robot tasks spanning varied failure scenarios.
- RoboFail includes 10 simulation tasks and 11 real-world tasks.
- The dataset’s simulation tasks are catalogued in Table 3.
- The dataset’s real-world tasks are catalogued in Table 4.
B.2 Comparison with GPT-3.5
REFLECT performs better with GPT-4 than GPT-3.5 for failure reasoning and correction planning, while GPT-4 more reliably processes irrelevant summary information.
- GPT-4 exceeds GPT-3.5 in failure reasoning and planning abilities.
- GPT-3.5’s failure localization accuracy decreases because it processes irrelevant summary information less effectively than GPT-4.
- The correction planning success rate decreases more significantly with GPT-3.5 than with GPT-4.
- Table 5 compares REFLECT’s performance with GPT-3.5.
- GPT-3.5 identifies only the failed bread placement, whereas GPT-4 identifies the incorrect action order and resulting placement error.
B.4 Human Evaluation
Human evaluation compares REFLECT-generated failure explanations with dataset ground-truth explanations using user scores for correctness and uncertainty.
- Two groups of three users compared ground-truth and REFLECT-generated explanations for each failure scenario.
- Users scored explanations 0 for incorrect, 1 for correct, and 2 when unsure.
- Failure scenarios were randomly shuffled in the questionnaires.