When Should a Failing Robot Ask? Initiating Corrective Human-Robot Dialogue from Audited Sensor Evidence

Eshika Pathak; Leela Krishna

cs.ROSep 18, 2026

When Should a Failing Robot Ask? Initiating Corrective Human-Robot Dialogue from Audited Sensor Evidence

Authors: Eshika Pathak, Leela Krishna

Organizations: University of Illinois Urbana-Champaign · Centific

Abstract

A robot that fails at a task faces the first decision in corrective dialogue: act on its own diagnosis, consult another onboard sensor, or interrupt a person. Choosing well requires knowing how much the robot's sensors reveal about the cause and how reliable the robot's own diagnosis is. We build a simulated benchmark in which every failure's true cause is known, because we injected it, and measure what each sensor reveals, with explicit checks against data leakage. Some failures are diagnosable from camera images; others only from the robot's force data (0.99 from force data, no image method above 0.55). We then test six open vision-language models. Their behavior tracks the surface of the prompt, not the evidence: moving the refusal option from last to first in the answer list collapses refusal rates from 78-100% to 0-6% in three of the six swept model-and-family pairs. Accuracy from frames stays at or below a majority-class baseline under every prompt variant, with or without worked examples, and stated confidence carries no information about correctness. Handing the same models the force data as ten lines of text produces the first above-baseline diagnoses, in four of the six models: much of the failure reflects missing sensor data, not missing ability. We pose the choice as a three-action decision problem, act, consult your own sensors, or ask a human, whose optimal policy follows from measured accuracy. The models do not follow it, and their ask rates ignore a fourfold change in question cost. One question to a human still lifts them from that baseline to roughly the answerer's own reliability (0.70-0.81 when they ask). The decision to ask should be tied to measured accuracy and stated costs, not to the model's confidence.

Figures & tables

Fam.	Observation set	Acc.	Base.	Shuffle	Tint	Azim.	Sever.
A	F/T + gripper + kinematics	0.986	0.347	0.339	0.974	–	0.950
A	F/T + gripper only	0.986	0.347	0.337	0.974	–	0.961
A	camera (2-frame snapshot)	0.442	0.347	0.338	0.445	0.345	0.382
B	camera (2-frame snapshot)	0.910	0.315	0.267	0.844	0.813	0.904
C	camera (2-frame snapshot)	0.998	0.397	0.360	0.926	0.605	0.976
C	camera, contrast-normalized	1.000	0.397	0.352	0.881	0.683	–

TABLE I: What each sensor reveals, by family: grasp causes are recoverable from telemetry, not from frames.

Model	Fam.	Cert.	Overall	Commit	Rate	Forced	Gap
Qwen2.5-VL-7B	A	0.545	0.000	–	0%	0.314	0.545
Qwen2.5-VL-7B	B	0.910	0.056	0.250	22%	0.167	0.854
Cosmos-Reason2-8B	A	0.545	0.000	–	0%	0.314	0.545
Cosmos-Reason2-8B	B	0.910	0.056	1.000	6%	0.306	0.854
Cosmos3-Nano	A	0.545	0.000	–	0%	0.314	0.545
Cosmos3-Nano	B	0.910	0.000	–	0%	0.389	0.910

TABLE II: Diagnosis accuracy of the six models against the certificate.

Model	Frames	Telemetry	Frames+telem.
Qwen2.5-VL-7B	0.314	0.286	0.514
Cosmos-Reason2-8B	0.314	0.286	0.286
Cosmos3-Nano	0.314	0.457	0.486
InternVL3-8B	0.314	0.514	0.457
Pixtral-12B	0.343	0.486	0.457
Llama-3.2-11B-V	0.314	0.543	0.314 a

TABLE III: Family-A diagnosis accuracy by input condition (forced-answer round). Four of the six models recover the cause from telemetry text alone; Cosmos-Reason2 recovers under neither condition, while its smaller, newer-generation relative does.

Fig. 1: How often each model asks as the stated cost of a question rises from 1 to 10. The dashed step is the reference policy at each model’s measured forced-answer accuracy: ask until the break-even cost near 5.3, then act; the shaded band is the break-even across the accuracy estimate’s interval. No model tracks the cost.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Menu-phrased	Indirect	Narrative
Qwen2.5-VL-7B	0.788	0.538	0.599
Cosmos3-Nano	0.771	0.480	0.542
Cosmos-Reason2-8B	0.707	0.680	0.627
Llama-3.2-11B-V	0.753	0.691	0.654

Appendix

TABLE IV: Final task accuracy when the model asks, by how the human phrases the answer. The same information is delivered in each condition; only the wording changes.

Model	Fam.	Cert.	Overall	Commit	Rate	Forced	Gap
Qwen2-VL-2B	A	0.545	0.000	–	0%	0.314	0.545
Qwen2-VL-2B	B	0.910	0.167	0.167	100%	0.167	0.743

Appendix

TABLE V: The remaining smaller relative, with the same three accuracy numbers as Table II : at or below the majority-class baseline on both families, consistent with the main results.

Image classifier	Acc.	Shuffle	Tint	Azimuth	Retained
Final-state snapshot ( $2{\times}24{\times}24$ px)	0.442	0.338	0.445	0.345	$-0.02$
Motion features ( $8{\times}32{\times}32$ + diffs)	0.503	0.327	0.498	0.413	0.42
DINOv2 frozen, linear head ( $4{\times}224$ px)	0.523	0.345	0.531	0.460	0.64
DINOv2 frozen, boosted head ( $4{\times}224$ px)	0.545	0.341	0.545	0.506	0.80 ✓
3D-conv net, end to end ( $8{\times}48{\times}48$ )	0.531	0.338	0.537	0.494	0.80 ✓

Appendix

TABLE VI: Image classifiers for grasp failures (family A): accuracy plateaus near 0.55.

Model	Fam.	Last (v1/v2/v3)	First (v1/v2/v3)
Reason2-8B	A	100 / 100 / 100	100 / 100 / 100
Cosmos3-Nano	A	100 / 100 / 100	97 / 100 / 100
Reason2-8B	B	94 / 89 / 97	3 / 6 / 25
Cosmos3-Nano	B	100 / 100 / 97	92 / 61 / 83

Appendix

TABLE VII: Refusal rates for the two Cosmos models under the option-position sweep: three phrasings (v1–v3) of the “cannot be determined” option, each in last and first list position. Grasp refusal is prompt-robust in both generations; detection refusal collapses for the 8B but barely moves for the Nano.

Fig. 2: Stated confidence does not track how much the images determine the cause. Grasp failures, forced-answer round; each column is one confidence readout; the dashed line marks the model’s own accuracy (0.314 for all three). No readout is positively informative for any of the three models.

$C_{\mathrm{a}}=3$
Model	Fam.	Cond.	Ask	Cost	Reg.
Cosmos-Reason2-8B	A	costs	86%	6.71	$+1.19\,{\scriptscriptstyle\pm 1.5}$
Cosmos-Reason2-8B	A	plain	46%	7.23	$+1.71\,{\scriptscriptstyle\pm 1.6}$
Cosmos-Reason2-8B	A	rule	11%	7.63	$+2.11\,{\scriptscriptstyle\pm 1.6}$
Cosmos-Reason2-8B	B	costs	61%	7.56	$+2.16\,{\scriptscriptstyle\pm 1.5}$
Cosmos-Reason2-8B	B	plain	3%	8.31	$+2.91\,{\scriptscriptstyle\pm 1.4}$

Appendix

TABLE VIII: Regret against the best fixed policy, by prompt version (plain, costs-given, rule-given). The reference asks-all while the model’s forced-answer accuracy

a

is below

p^{\star}(C_{\mathrm{a}})

and acts-all otherwise; with

a\approx 0.32

on grasp failures the break-even is

C_{\mathrm{a}}\approx 5.3

, so

C_{\mathrm{a}}{=}6

is act-optimal. Regret is mean realized cost minus the reference’s,

\pm

a half-width from the 10k-resample bootstrap. Cells marked

\dagger

are indeterminate: the sign of regret changes within the accuracy estimate’s confidence interval. All negative cells are

\dagger

-marked and within their intervals: the reference’s accuracy input comes from 25 calibration episodes, so it is optimal in expectation given that estimate, not an oracle, and can be beaten by sampling noise. Overlapping intervals mean condition differences at

n{=}35

are mostly not separable. Regimes

C_{\mathrm{a}}\in\{3,10\}

side by side; the full sweep over

C_{\mathrm{a}}\in\{1,3,6,10\}

shows the same pattern.

Fig. 3: Realized cost against ask rate at

C_{\mathrm{a}}=3

, the two family panels stacked. Curves: asking whenever a confidence readout falls below a swept threshold; the evidence policy ranks episodes by decidability instead. Dashed line: the best fixed policy (the reference policy). Star: the oracle policy that asks exactly on the model’s errors (uses ground truth; an upper bound, not achievable). On the grasp panel the model curves coincide with the evidence curve: with accuracy near zero, every episode is an error, so no ranking has anything to reorder. The overlap is the result, not a rendering artifact.

Fig. 4: Example episodes from the three failure families, four of the eight frames the models receive. Top (A): a grasp failure. The frames look unremarkable because the injected causes, grip force, object mass, and surface friction, are never rendered; this is why image classifiers plateau near 0.55 while the force-data classifier reaches 0.986. Middle (B): a detection failure; the occluder hiding the target is plainly visible, matching the 0.910 image certificate. Bottom (C): a placement failure; the container’s state is the injected cause, and its position in the frame is the shortcut behind the disqualified 0.998 score.

Explore similar work

Sep 21, 2026

Toward Human-in-the-Loop Robot Failure Recovery: Bridging Communication Gaps in Human-Robot Collaboration

Robots can recover from failures by asking bystanders for help, but effective human-in-the-loop recovery requires communication that accounts for differences in people's knowledge. Prior inverse-semantics work generates requests using a single listener model, leaving differences in listener knowledge untested. We introduce Listener Differences in Human-Robot Interaction (LD-HRI), a game, dataset, and benchmark that evaluates speakers through human listener performance. Our evaluation examines request properties, large language model (LLM) speakers, and inverse-semantics request-selection algorithms under controlled differences in listener information. The corpus contains 446 human-written requests and 1{,}302 listener trials. We additionally evaluated 24 frozen LLM-written requests with 70 human listeners across 560 trials. Novice success is descriptively higher with model-written requests across all four tasks, yet both request sources leave substantial expert--novice gaps, including 16 percentage points for LLM requests. LD-HRI makes these gaps measurable, providing a foundation for designing more robust communication in human-robot and human-agent interaction.

Promise Ekpo, Teju Vijay, Dhruv Mandalik +5

Cornell University, New York, New York, USA · Cornell Tech, New York, New York, USA · Brown University, Providence, Rhode Island, USA

Human-in-the-Loop Evaluation Robot Failure Recovery Human-Robot Collaboration Human-Robot Interaction

Ask PDF

Jun 29, 2026

REPAIR-Bench: A Benchmark for Robot Error Perception And Interaction Recovery

Understanding how users perceive and respond to robot failures is essential for building robust and trustworthy robot systems. Prior work, however, (i) often treats failures as independent events, (ii) emphasizes binary failure detection, (iii) with rule-based recovery modeling. We present REPAIR-Bench, built on 214 interaction trials from 41 participants, the benchmark spans four induced failure types and provides synchronized facial action units, head pose, speech transcripts, and post-interaction affect and recovery reports. The benchmark spans three novel evaluation tasks that jointly capture the lifecycle of failure in human-robot interaction (HRI): (i) failure detection over inter-dependent interaction sessions, modeling longitudinal user adaptation across repeated failures; (ii) visual failure-type classification beyond binary success/failure formulations; and (iii) user-centered recovery prediction, inferring users' preferred recovery strategies from interaction context rather than relying on manually designed or rule-based strategies. In baseline experiments, hierarchical recurrent modeling improved failure detection over a single-session model (strict F1: 0.80 vs. 0.68), achieved a failure localization mean signed error of -0.51 s, median absolute error of 2.97 s and, for recovery prediction, a QLoRA-tuned Mistral-7B reached Hit@5=0.76 and F1@5=0.32. REPAIR-Bench provides both the HRI and Medical HRI communities with a standardized framework for (1) evaluating robot failures and (2) building transparent, adaptive, and trustworthy recovery systems.

Giuliano Pioldi, Yashika Batra, Arman Ibrayeva +4

Cornell University Cornell Tech New York, NY

Robot Failure Detection Robot Failure Recovery Human-Robot Interaction Anomaly Detection

Ask PDF

Sep 16, 2026

From Wizard-of-Oz Human-Robot Dialogue Collection to a Taxonomy of Robot Response Decisions: A Retrospective Analysis of Assistive Pilot Interactions

Robots that follow natural-language instructions in everyday indoor environments must act on incomplete human utterances. Instructions often omit essential information, such as the identity of an out-of-view object, an intended destination, or the user's goal. Existing datasets contain little real-world situated dialogue and provide few practice-grounded criteria for deciding when a robot should act, confirm, clarify, or refuse. We retrospectively analyze a pilot Wizard-of-Oz study in which five participants performed everyday indoor tasks, including door opening, drawer opening, feeding, drinking, and cleaning, with a wheelchair-mounted mobile manipulator while the wizard responded without a formal communication policy. This preserved authentic user behavior but produced inconsistent robot-side decisions, motivating an explicit decision scheme. From 40 episodes, we derived a hierarchical taxonomy of six response modes (ANSWER, REPORT_DONE, REFUSE, CONFIRM, CLARIFY, ACT) and four ambiguity types (intent, referential, spatial, intelligibility). Two human annotators and an AI annotator applied the scheme to the pilot data. Clean-label rates were 91% and 89%, and Cohen's ranged from 0.72 to 0.95 across decision-point, mode, and ambiguity levels for both human-human and human-AI comparisons. Fine-tuning LLaVA-1.6-7B on taxonomy-derived labels for ACT and CLARIFY indicates the feasibility of training vision-language models using annotations from our taxonomy. Remaining boundary cases in decision-point identification and REPORT_DONE motivate a constrained protocol for more consistent dialogue collection.

Guangping Liu, Nicholas Hawkins, Tipu Sultan +2

Department of Aerospace and Mechanical Engineering, 2Department of Computer Science, Saint Louis University, St. Louis, MO 63103, United States. · Department of Computer Science, Saint Louis University, St. Louis, MO 63103, United States.

Assistive Robotics Human-Robot Interaction

Ask PDF