Enhancing Joint Human-AI Inference in Robot Missions: A Confidence-Based Approach

Clara Colombatto; Duc-An Nguyen; Ingmar Posner; Nick Hawes; Raunak Bhattacharyya; Steve Fleming

REVIEW 6 major objections 6 minor 4 references

Enhancing Joint Human-AI Inference in Robot Missions: A Confidence-Based Approach

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Selecting the higher-confidence inference between a human and an AI decision-support system raises team accuracy in a simulated robot teleoperation task, provided the AI's confidence is well calibrated.

desk verdict A well-run behavioral study of confidence-based arbitration in human-AI teams, but the headline statistics need re-analysis and the synthetic AI-DSS limits external claims. read the letter →

arxiv 2508.03293 v1 pith:JAYMOV73 submitted 2025-08-05 cs.HC cs.RO

Duc-An Nguyen , Clara Colombatto , Steve Fleming , Ingmar Posner , Nick Hawes , Raunak Bhattacharyya This is my paper

classification cs.HCcs.RO

keywords MaximumConfidenceSlatingjointhuman-AIinferencecalibrationrobotteleoperationdecisionsupportsystemhuman-robotteamingmetacognitionAI-assistedmaking

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that when a human and an AI decision-support system disagree, the team should simply adopt the inference backed by the higher confidence. In a simulated robot teleoperation task where operators infer which of two robots has a smaller control delay, this maximum-confidence heuristic produced higher joint accuracy than letting the human make the final call or letting a Thompson sampling bandit arbitrate. The advantage, however, is not unconditional: pairing humans with a well-calibrated AI (one whose confidence tracks its correctness) gave roughly 75% accuracy, while a poorly calibrated AI gave about 70% and could even push the team below individual human performance. The authors present this as the first demonstration of maximum-confidence slating for human-AI joint inference in a robotics task that combines latent variables, temporal observation, and action-based inference.

What carries the argument

The central object is the maximum-confidence slating (MCS) heuristic: given two agents' probabilistic predictions, accept the classification of the agent whose confidence (distance of the prediction from 0.5) is larger. The study's machinery also includes a synthetic AI-DSS whose binary inference accuracy is fixed at 70% and whose confidence is sampled from distributions built from separately collected human data, split into well-calibrated (AUROC2 ≥ 0.65) and poorly calibrated (AUROC2 ≤ 0.55) variants. AUROC2, the area under the receiver-operating characteristic for confidence versus correctness, is used to measure calibration. MCS is the decision rule that carries the argument; the synthetic AI-DSS and calibration split are what allow the authors to isolate the effect of calibration on team accuracy.

What would settle it

Reproduce the two-robot teleoperation study with a real trained classifier as the AI-DSS (e.g., a network trained on the recorded trajectories) and measure whether MCS still beats human-initiated inference when the AI's AUROC2 is well calibrated; alternatively, sweep the fixed AI accuracy from 55% to 95% in the synthetic design and check whether the MCS advantage over human-initiative inference persists or reverses.

Watch

Extended reading notes

Core claim

The central claim is that maximum-confidence slating (MCS) — taking the inference of whichever agent, human or AI, reports the higher confidence — yields better team accuracy than either agent alone when the AI's confidence is well calibrated. In a between-subject study with 100 participants (80 retained), the MCS-based joint inference outperformed human-initiated inference after seeing the AI recommendation ($t(4828)=8.52$, $p<0.001$) and outperformed a Thompson sampling bandit that learned which agent to trust ($t(4828)=4.74$, $p<0.001$). The benefit was modulated by the AI's calibration: with a well-calibrated AI the team reached $74.95\%\pm0.68\%$ accuracy, versus $70.47\%\pm0.65\%$ with a poorly calibrated AI. The authors further show that poorly calibrated AI undermines MCS because the AI's overconfident wrong inferences override correct human judgments, and that humans make more effective inference revisions when the AI is well calibrated.

Load-bearing premise

The AI-DSS is not a real AI but a synthetic simulator whose binary accuracy is fixed at 70% and whose confidence values are drawn from earlier human responses, so if actual deployed AI confidence behaves differently, the measured MCS advantage and calibration effect may not transfer.

Editorial extensions

If this is right

If MCS works as claimed, robot teleoperation teams can be improved without retraining models: just ask both agents for confidence and arbitrate by it.
The dependence on calibration implies that AI-DSS for joint inference should be evaluated on metacognitive sensitivity (AUROC2), not just accuracy, before deployment.
Well-calibrated AI not only improves the joint decision but also steers human revisions toward correct answers, so calibration has a training-like effect on the human partner.
The outperformance over the Thompson sampling bandit suggests that trial-by-trial confidence is a better arbiter than learned overall trust rates, at least in this task.
Poorly calibrated AI can be worse than no AI, so confidence-based joint inference should be gated on calibration quality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The fixed 70% AI accuracy likely matters: if real AI accuracy were much higher or lower, the relative advantage of MCS over human-initiative arbitration could shrink or reverse, since the high-confidence agent would be wrong more often; a natural extension is to sweep AI accuracy across levels.
The results were obtained with a synthetic AI whose confidence is drawn from human data, independent of the particular human's errors; real AI-DSS confidence may be correlated with human mistakes on hard trials, which would reduce the independence MCS relies on.
The four-point Likert confidence scale and binary two-robot choice are simple; mapping MCS to continuous confidence and multi-class or sequential decisions is a natural next test.
The finding that humans revise inferences more positively with a well-calibrated AI suggests an intervention: displaying calibration or metacognitive-sensitivity information alongside confidence may shift human trust in the right direction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this paper before reading it. First, it's a well-designed behavioral study showing that when a human and an AI-DSS each give a binary inference plus a confidence rating, picking the more confident one (MCS) beats letting the human decide, and beats a Thompson-sampling bandit, in a simulated robot delay-discrimination task. Second, the "AI-DSS" is not a real AI system: it's a sampler that draws confidence values from distributions of human participants who did the same task, with accuracy fixed at 70%. That makes the calibration manipulation clean but limits what we can say about real learned AI systems.

What's new: it extends MCS from human-human dyads to human-AI teams in a task that combines latent variables, temporal observation, and action—the HRI-relevant combination that static oddball tasks lack. The human inference-revision data (Fig. 4) are the most solid part: well-calibrated AI leads to more incorrect-to-correct changes and fewer correct-to-incorrect reversals. That's a useful, believable finding.

Where it gets soft. The joint-accuracy headline relies on trial-level t-tests with degrees of freedom in the thousands (e.g., t(4828)=8.52). Those treat each trial as independent, ignoring participant clustering, so the p-values are far too small. A mixed-effects analysis or participant-level bootstrap is needed. Also, the paper excludes 20% of participants based on a 65% accuracy threshold; that's pre-registration-looking but not pre-registered, and it can bias performance upward. There's also an internal numerical inconsistency: MCS with well-calibrated AI is reported as 64.28% in one sentence and 74.95% in the next for what looks like the same condition. And in Fig. 8, the text reports t(1716)=-1.15 with p<0.001, which can't be right for a two-sided test.

The calibration effect is engineered by construction—well vs poorly calibrated AI are defined by sampling from human subgroups with AUROC2 ≥0.65 vs ≤0.55. That's a legitimate manipulation for a behavioral study, but it means the "AI" has no learned representation of difficulty; real AI confidence is typically correlated with input difficulty in ways this sampler doesn't capture. The paper acknowledges the synthetic construction in the limitations, which is honest, but the main results still stand on it.

Who should read this: anyone designing confidence-based arbitration for human-robot teams, and researchers studying metacognition in joint decision-making. It's a worthwhile contribution that needs statistical re-analysis and clearer reporting, not desk rejection.

My recommendation: send it to peer review, but the reviewers should push for participant-level analysis and a clear resolution of the reported accuracy inconsistency.

Referee Report

6 major / 6 minor

Summary. The paper proposes applying Maximum Confidence Slating (MCS) to joint human-AI inference in a simulated robot teleoperation task. In each trial, participants teleoperate two robots, infer which has the lower control delay, report a confidence rating, then see an AI decision-support system (AI-DSS) inference with a confidence rating and make a final inference; the team decision is then selected by MCS. The AI-DSS is synthetic: it samples binary inferences with a fixed 70% accuracy and samples confidence ratings from distributions of human confidence obtained in a prior data collection, with well- and poorly-calibrated variants constructed by selecting human subgroups with AUROC2 above 0.65 or below 0.55. The study reports that MCS outperforms human-initiated inference and a Thompson sampling bandit, that well-calibrated AI-DSS improves MCS accuracy relative to poorly calibrated AI-DSS, and that human inference revisions depend on AI-DSS calibration. The authors acknowledge the synthetic AI-DSS and other limitations in the Discussion.

Significance. If the results hold, the paper would provide a simple, practically relevant decision rule for human-AI teams in robot teleoperation and would extend MCS from human-human to human-AI settings. The study has notable strengths: a sizeable sample (N=100, 80 after exclusion), a representative task combining latent variables, temporal processes, and action-based inference, and a clever experimental manipulation of AI confidence calibration. The behavioral findings on human revision dynamics are interesting in their own right. However, the central comparisons rest on a synthetic AI-DSS rather than a real AI model, the statistical reporting appears to violate trial independence, and there are internal inconsistencies in the main numerical results. These issues are load-bearing for the headline claims, so the paper needs substantive revision.

major comments (6)

[AI Decision Support System (AI-DSS)] The AI-DSS is not a learned model but a synthetic confidence sampler whose binary inference accuracy is fixed at 70% and whose confidence is drawn from distributions of human confidence in a separate data collection. The abstract and Results present the headline comparisons (e.g., t(4828)=8.52 for MCS vs human-initiated; 74.95% vs 70.47% for well- vs poorly-calibrated AI-DSS) as evidence about 'joint human-AI inference.' Since the AI-DSS has no input-dependent prediction mechanism, the experiment effectively studies MCS with a simulated 'second human' rather than with an AI system; real AI-DSS confidence is typically correlated with input difficulty and with human error structure in ways not captured by the marginal distributions used here. The Discussion acknowledges this, but the central claim's external validity remains a load-bearing gap that should be addressed in the title, abstract, and conclusions, or supplemented with an experiment using an actual trained classifier.
[Results/Joint Inference Performance] The main t-tests use trial-level degrees of freedom, e.g., t(4828)=8.52 for MCS vs human-initiated inference and t(4828)=4.74 for MCS vs Thompson sampling. Because each participant contributes up to 100 trials, these observations are not independent; treating them as independent units inflates the effective sample size and can produce artificially small p-values. The authors should reanalyze the data with models that account for participant-level clustering (e.g., mixed-effects logistic regression with random intercepts for participants) or with participant-level summary statistics, and report the resulting effect sizes and confidence intervals for the central claims.
[Results/Joint Inference Performance] The same condition is reported with two different accuracies in the paragraph describing Fig. 7: 'the MCS-based joint inference achieved an accuracy of (64.28% ± 0.59%)' with a well-calibrated AI-DSS, and two sentences later 'MCS-based joint inference with a well-calibrated AI-DSS yielded (74.95% ± 0.68%) accuracy.' This internal contradiction concerns the central numerical result and must be resolved; the authors need to clarify which number corresponds to Fig. 7, correct the text or figure, and re-run the associated significance tests.
[Results/Joint Inference Performance, Fig. 8] The sentence 'HP slightly outperformed MCS (t(1716) = −1.15, p <0.001)' is internally inconsistent: a two-sided t statistic of −1.15 with 1716 degrees of freedom would yield a large p-value, not p<0.001, and the sign of the statistic does not match the stated direction of the comparison. This error affects the evidence for the claimed limitation of MCS with poorly calibrated AI-DSS and needs to be corrected and re-analyzed.
[AI Decision Support System (AI-DSS)] The well- vs poorly-calibrated AI-DSS conditions are constructed by sampling confidence from participants with AUROC2 ≥ 0.65 versus ≤ 0.55, while fixing AI-DSS accuracy at 70%. Because MCS selects the agent with higher confidence, and high AUROC2 by definition means that higher confidence is more often assigned to correct than incorrect inferences, the result that MCS performs better with the well-calibrated AI-DSS is partly guaranteed by the construction of the manipulation. To establish that calibration has an additional, non-trivial effect on team inference, the authors should provide a simulation or analytic calculation using the constructed distributions that shows the expected MCS accuracy under each condition; otherwise the empirical contrast risks being a restatement of the selection rule.
[Experimental Procedure] The post hoc exclusion of participants with accuracy below 65% (20 of the 100 participants) can bias performance estimates upward, and the main comparisons in Figs. 6 and 7 are based on the remaining 80 participants. The authors should report exclusion counts separately for the well- and poorly-calibrated conditions, justify that the threshold did not differentially affect the two arms, and provide a robustness analysis including all participants or using a pre-specified exclusion criterion.

minor comments (6)

[Results/Joint Inference Performance] In the paragraph preceding Fig. 8, 'virual pairing' should be 'virtual pairing.'
[Results, Fig. 5] The reported 'r2 = 0.183' appears to be R² from a linear regression; please use consistent notation (R² or r²) and report the regression coefficient and confidence interval.
[Methodology, AI Decision Support System] The term 'confidence calibration' is used for AUROC2, but AUROC2 measures metacognitive sensitivity/discrimination, not calibration in the sense of predicted probability matching observed frequency. Consider using 'metacognitive sensitivity' or clearly defining the term.
[Problem Statement] MCS is defined using distance from 0.5 for probabilistic predictions, but the experimental implementation uses a four-point Likert confidence scale. Please clarify how the continuous definition maps to the discrete ratings, and how ties are resolved.
[Results/Joint Inference Performance] It is not fully clear whether 'Human Individual Inference' and 'Joint Inference by Human' are computed on all trials or only on the subset of trials where the human and AI-DSS gave opposing inferences; please specify the exact trial sets for each comparison in the figure captions or text.
[Results/Joint Inference Performance] The Thompson sampling bandit baseline is not described in enough detail (e.g., priors, update schedule, whether it is trained online per trial). Please provide the algorithm details or a reference so the comparison is reproducible.

Circularity Check

1 steps flagged · score 6.0 of 10

Calibration-dependent MCS benefit is imposed by construction; the MCS-vs-human/TS comparison retains independent empirical content.

self definitional [Methodology, 'AI Decision Support System (AI-DSS)' section; Results, 'Joint Inference Performance'.]
"We use AUROC2 as a measure of confidence calibration... An AUROC2 of 0.85 means there’s an 85% chance that a person gives higher confidence to correct answers than to incorrect ones... Based on this, we divided the dataset into two groups for AI-DSS learning: the well-calibrated AI-DSS samples from the data of humans with well-calibrated confidence (AUROC2 ≥ 0.65). The poorly calibrated AI-DSS sample from the data of participants with poorly-calibrated confidence (AUROC2 ≤ 0.55)."

The headline calibration result — MCS with well-calibrated AI-DSS at 74.95% vs 70.47% with poorly-calibrated AI-DSS — is a consequence of how the two AI-DSSs are constructed. AUROC2 is by definition the probability that a higher confidence is assigned to correct than to incorrect inferences, and MCS selects the inference with higher confidence. Sampling AI-DSS confidence from humans with AUROC2 ≥ 0.65 versus ≤ 0.55 therefore guarantees that the high-confidence AI inference is more often correct in the 'well-calibrated' arm. The observed calibration effect is built into the independent variable rather than discovered from human-AI interaction.

full rationale

The study's central comparison — MCS-based joint inference outperforming human-initiated inference and Thompson-sampling bandits (t(4828)=8.52 and t(4828)=4.74) — is not circular by construction: it depends on measured human confidence, human inference revision, and the synthetic AI-DSS's fixed 70% accuracy, and could in principle have gone the other way. However, the abstract and results also claim that the benefit is 'regulated by the confidence calibration of the AI agent,' and this particular claim reduces to the construction. The well-calibrated and poorly-calibrated AI-DSS are defined by selecting human subgroups with high vs low AUROC2, and AUROC2 is exactly the property MCS exploits (higher confidence should mean more likely correct). Thus the 74.95% vs 70.47% calibration contrast is a near-tautological outcome of the independent-variable definition, not an empirical discovery about real AI-DSS behavior. Self-citations, such as Nguyen et al. 2025 for robot-mission MCS, are not load-bearing here because the MCS background is independently supported by Bahrami et al. 2010 and Koriat et al. 2012. The paper explicitly acknowledges the synthetic AI-DSS in its limitations, but the headline calibration effect remains definitionally tied to the AUROC2-based construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claims rest on the synthetic AI-DSS construction, the comparability of human and AI confidence scales, and post-hoc participant exclusions. The free parameters are experimental design choices, not physical or mathematical constants.

free parameters (5)

AI-DSS accuracy = 70%
The AI-DSS's binary inference is fixed at 70% accuracy to simulate an imperfect decision support system; this value enters every joint-accuracy comparison.
Well-calibrated AI-DSS threshold = AUROC2 >= 0.65
Defines the well-calibrated condition from which the AI-DSS confidence is sampled; changing this cut changes the calibration manipulation.
Poorly-calibrated AI-DSS threshold = AUROC2 <= 0.55
Defines the poorly-calibrated condition; paired with the >=0.65 cut it creates the between-subject contrast.
Participant exclusion thresholds = accuracy < 65%; same confidence > 95/100 trials
Post-hoc criteria that removed 20 of 100 participants; affects the sample and can inflate the accuracy estimates.
Staircase delay parameters = 20 ms step, 35 ms start, levels 20-100 ms
Hand-set parameters of the adaptive staircase that controls task difficulty and target ~70% human accuracy.

assumptions (4)

domain assumption Human and AI-DSS receive the same information during inference.
Stated in the Limitations; if information sources differ in real missions, confidence may not be comparable and MCS may not hold.
domain assumption Human and AI-DSS confidence values are commensurable on the same 4-point Likert scale.
MCS requires direct comparison of confidence numbers; the paper assumes the human's 1-4 rating and the AI-DSS's sampled rating are on the same metric.
domain assumption Each agent's confidence is (at least weakly) calibrated to accuracy, so higher confidence indicates higher probability of being correct.
This is the operating premise of MCS; the paper measures AUROC2 for humans and engineers it for the AI-DSS.
domain assumption The adaptive staircase keeps accuracy near 70% and prevents ceiling or floor effects.
The delay-difference staircase assumes the 20-100 ms range spans the participant sensitivity range; if not, accuracy estimates and MCS comparisons are distorted.

invented entities (1)

Synthetic AI-DSS
purpose: Simulates an AI decision support system by sampling confidence from separately collected human confidence distributions conditioned on task difficulty, with binary inference accuracy fixed at 70%.
The AI-DSS is not a trained model that processes the task inputs; it is an experimental confederate built from human data. Its confidence calibration is engineered, so the study provides no falsifiable handle on how a real AI-DSS would behave outside this setup.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Joint Human-AI Inference in Robot Missions: A Confidence-Based Approach." pith.science (2026). https://pith.science/paper/JAYMOV73

@misc{pith2026250803293,
  author       = {Pith},
  title        = {Pith review of: Enhancing Joint Human-AI Inference in Robot Missions: A Confidence-Based Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JAYMOV73}},
  note         = {Machine review of arXiv:2508.03293}
}

read the original abstract

Joint human-AI inference holds immense potential to improve outcomes in human-supervised robot missions. Current day missions are generally in the AI-assisted setting, where the human operator makes the final inference based on the AI recommendation. However, due to failures in human judgement on when to accept or reject the AI recommendation, complementarity is rarely achieved. We investigate joint human-AI inference where the inference made with higher confidence is selected. Through a user study with N=100 participants on a representative simulated robot teleoperation task, specifically studying the inference of robots' control delays we show that: a) Joint inference accuracy is higher and its extent is regulated by the confidence calibration of the AI agent, and b) Humans change their inferences based on AI recommendations and the extent and direction of this change is also regulated by the confidence calibration of the AI agent. Interestingly, our results show that pairing poorly-calibrated AI-DSS with humans hurts performance instead of helping the team, reiterating the need for AI-based decision support systems with good metacognitive sensitivity. To the best of our knowledge, our study presents the first application of a maximum-confidence-based heuristic for joint human-AI inference within a simulated robot teleoperation task.

Figures

Figures reproduced from arXiv: 2508.03293 by the authors.

**Figure 2.** Histograms of human confidence levels associated [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Study Interface: After teleoperating the two robots, participants make an initial inference and provide associated confidence (left panel). Then, the AI-DSS provides its inference and confidence (middle panel). If the participant decides to change their initial inference, they provide an updated inference and associated confidence (right panel). The joint human-AI inference is then decided using MCS (top right). Neg… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Dynamics of human inferences: AI-DSS confidence-calibration influences whether humans change from incorrect to correct inferences. made correct inferences but changed to incorrect inferences after observing the AI-DSS inference and its confidence. The number of each ty…

**Figure 7.** Figure 7: Influence of varying AI type on team inference [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Inference accuracy obtained by pairing human par [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[3]

ACM computing surveys, 56(7): 1–32

Who’s in charge here? A survey on Trustworthy AI in Variable Autonomy Robotic Systems. ACM computing surveys, 56(7): 1–32. Mozannar, H.; and Sontag, D. 2020. Consistent estimators for learning to defer to an expert. In International Confer- ence on Machine Learning (ICML). Natarajan, S.; Mathur, S.; Sidheekh, S.; Stammer, W.; and Kersting, K. 2025. Human-...

work page arXiv 2020
[4]

Human-AI Symbiosis: A Survey of Current Approaches

Meaningful human control and variable autonomy in human-robot teams for firefighting. Frontiers in Robotics and AI, 11: 1323980. Wang, G.; Wang, W.; Ding, P.; Liu, Y .; Wang, H.; Fan, Z.; Bai, H.; Hongbiao, Z.; and Du, Z. 2023. Development of a search and rescue robot system for the underground building environment. Journal of field robotics, 40(3): 655–6...

work page Pith review arXiv 2023
[2023]

In ACM Conference on Fairness, Accountability, and Transparency (FAccT)

Towards a Science of Human-AI Decision Making: An Overview of Design Space in Empirical Human-Subject Studies. In ACM Conference on Fairness, Accountability, and Transparency (FAccT). Ma, S.; Wang, X.; Lei, Y .; Shi, C.; Yin, M.; and Ma, X

work page
[2024]

In Conference on Human Factors in Computing Systems (CHI)

Are you really sure? Understanding the effects of hu- man self-confidence calibration in AI-assisted decision mak- ing. In Conference on Human Factors in Computing Systems (CHI). Madras, D.; Pitassi, T.; and Zemel, R. 2018. Predict re- sponsibly: improving fairness and accuracy by learning to defer. Advances in Neural Information Processing Systems (NeurI...

work page 2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [3]

ACM computing surveys, 56(7): 1–32

Who’s in charge here? A survey on Trustworthy AI in Variable Autonomy Robotic Systems. ACM computing surveys, 56(7): 1–32. Mozannar, H.; and Sontag, D. 2020. Consistent estimators for learning to defer to an expert. In International Confer- ence on Machine Learning (ICML). Natarajan, S.; Mathur, S.; Sidheekh, S.; Stammer, W.; and Kersting, K. 2025. Human-...

work page arXiv 2020

[2] [4]

Human-AI Symbiosis: A Survey of Current Approaches

Meaningful human control and variable autonomy in human-robot teams for firefighting. Frontiers in Robotics and AI, 11: 1323980. Wang, G.; Wang, W.; Ding, P.; Liu, Y .; Wang, H.; Fan, Z.; Bai, H.; Hongbiao, Z.; and Du, Z. 2023. Development of a search and rescue robot system for the underground building environment. Journal of field robotics, 40(3): 655–6...

work page Pith review arXiv 2023

[3] [2023]

In ACM Conference on Fairness, Accountability, and Transparency (FAccT)

Towards a Science of Human-AI Decision Making: An Overview of Design Space in Empirical Human-Subject Studies. In ACM Conference on Fairness, Accountability, and Transparency (FAccT). Ma, S.; Wang, X.; Lei, Y .; Shi, C.; Yin, M.; and Ma, X

work page

[4] [2024]

In Conference on Human Factors in Computing Systems (CHI)

Are you really sure? Understanding the effects of hu- man self-confidence calibration in AI-assisted decision mak- ing. In Conference on Human Factors in Computing Systems (CHI). Madras, D.; Pitassi, T.; and Zemel, R. 2018. Predict re- sponsibly: improving fairness and accuracy by learning to defer. Advances in Neural Information Processing Systems (NeurI...

work page 2018