Pith. sign in

REVIEW 4 major objections 5 minor 21 references

A robot can shape a person's belief across repeated encounters to keep its long-term influence without sacrificing trust, by adaptively switching between influence and trust modes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 03:24 UTC pith:ELZHV4PH

load-bearing objection Solid system paper with real user-study results, but the belief-tracking validation is circular and the adaptive arbitration's belief channel needs an ablation to be credible. the 4 major comments →

arxiv 2607.25327 v2 pith:ELZHV4PH submitted 2026-07-28 cs.RO

Belief-Aware Influence and Trust (BAIT): Shaping Human Belief During Repeated Human-Robot Interaction

Ye-Ji Mun , Mahsa Golchoubian , Shahabedin Sagheb , Yan Bai , Tianhao Ji , Dylan P. Losey , Katherine Driggs-Campbell This is my paper
classification cs.RO
keywords long-term human-robot interactiontrustinfluencebelief-aware planninghierarchical particle filtermodel predictive path integralrepeated lane mergingautonomous vehicle navigation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a robot facing the same person over many encounters can actively steer that person's evolving perception of the robot, rather than treating each encounter as isolated. The proposed controller, BAIT, tracks a two-timescale model of the human — a fast-changing driving strategy and a slower belief about whether the robot is assertive or defensive — and plans actions that explicitly balance long-term influence against trust, with immediate task performance enforced as a hard safety constraint. In simulation, a 30-person user study, and a real-world vehicle deployment, BAIT is reported to match the task performance of controllers that maintain influence through unpredictability while producing higher user trust, and to run online at about 0.11 seconds per step. If these results hold, the paper opens a practical route to long-term human-robot cooperation in which robots do not have to choose between being effective and being trusted.

Core claim

The central claim is that the influence–trust trade-off in repeated human-robot interaction can be managed online by separating the human's latent state into a fast strategy variable and a slow belief variable, inferring both with a hierarchical particle filter, and then planning with a belief-aware sampling-based controller that propagates posterior moments deterministically. The controller arbitrates between an influence cost, which promotes unpredictable entropy-increasing behavior to keep the robot from being exploited, and a trust cost, which promotes transparent predictable legibility, while a conditional value-at-risk constraint blocks any trajectory that risks the immediate task. An

What carries the argument

The central object is the two-timescale latent hierarchy: a slow long-term belief φ∈[0,1] (assertive-to-defensive perception of the robot) that shapes a fast short-term strategy z (the human's weighting of progress versus safety), together with the conditional coupling that makes the belief observable only through the strategy. This hierarchy is tracked by a two-layered particle filter — long-term particles drift through a mixture Beta kernel and are updated by both action evidence and a geometric yield-belief proxy, while short-term particles mean-revert through an Ornstein-Uhlenbeck process and are reweighted by Boltzmann action likelihoods. For planning, the full mixed-observability belie

Load-bearing premise

The central result collapses if a real human driver's belief does not evolve through the assumed two-timescale structure — a slow perceptual belief that only acts through a fast strategy — because the filter, the planner's rollout surrogate, and the simulated humans all share that same model family.

What would settle it

In a repeated lane-merging study, collect a per-trial measure of the human's belief that the robot will yield (for example, a short post-merge rating) and compare it with the filter's inferred belief. If the controller achieves high task performance while its inferred belief tracks the reported belief no better than a constant prior, the claim that the robot is actually shaping human belief would be falsified. Alternatively, run the adaptive controller with the belief channel artificially frozen while preserving the same observable actions; if trust ratings are unchanged, the belief-shaping me

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Repeated human-robot interaction can be planned in real time: the deterministic surrogate avoids the exponential branching of full belief-tree solvers, with reported control at about 0.11 seconds per step.
  • If the claim is right, adaptive mode switching lets a robot keep its long-term influence (preventing humans from settling into exploitative behavior) while preserving significantly more user trust than pure influence controllers.
  • Transparent, trust-maximizing robots are exploitable over repeated interactions: both the stationary baseline and the pure-trust variant saw human lane progress rise over episodes as drivers learned the robot would yield.
  • The belief-shaping effect transfers from simulation to a 30-person user study and to physical vehicle deployments, so the mechanism is not limited to one environment.
  • Immediate task performance can be made non-negotiable: the CVaR constraint filters out belief-shaping trajectories that jeopardize the merge or collision avoidance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the paper reports only binary mode switching, a natural testable extension is continuous arbitration over β; if the trust–influence frontier is well behaved, a smoother schedule could dominate the binary hysteresis rule.
  • The geometric yield-belief proxy is a stand-in for real perception; replacing it with elicited self-reports of expected robot yielding would test whether the filter is tracking beliefs or just kinematics.
  • A stress test of the model family: run the same controller against simulated humans whose belief-update kernel differs from the assumed Beta/OU structure (for example, a belief that updates directly from robot actions rather than only through the strategy variable); if trust and task performance degrade sharply, the result may owe more to model fit than to the arbitration logic.
  • The user study limits generalization to a single scenario and 20 episodes per condition; whether the trust benefit persists across heterogeneous road users or longer exposure remains open.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes BAIT, a controller for repeated human-robot interaction that tracks a two-timescale latent state — a fast strategy variable z_t and a slow belief variable phi_t — with a hierarchical particle filter, and uses a deterministic-surrogate MPPI planner to explicitly trade off long-horizon influence against human trust while enforcing task performance as a CVaR constraint. The authors evaluate BAIT-trust, BAIT-infl, and an adaptive BAIT-adapt variant in a CARLO-based simulated lane-merging task, in a 30-participant in-person user study, and on Polaris GEM vehicles. The central claim is that adaptive arbitration between influence and trust achieves cumulative task performance comparable to influence-only baselines while preserving significantly higher user trust. The paper also claims real-time scalability (~0.11 s/step) by replacing full MOMDP belief-tree evaluation with a posterior-moment surrogate.

Significance. If the belief-estimation channel were independently validated, BAIT would be a meaningful advance: it provides a concrete, real-time architecture for the influence-trust trade-off in repeated HRI, with explicit equations, a user study, and a hardware demonstration. The deterministic-surrogate MPPI planner is a practical contribution over prior MOMDP-based approaches, and the paper makes falsifiable quantitative predictions (e.g., lane-progress differences across modes). However, the current validation of the belief channel is internally circular, and the adaptive controller is never compared against a version that removes the belief-based switching signal. As a result, the distinctive 'belief-aware' mechanism is not yet supported as the cause of the reported trust/task outcomes. These gaps are fixable in revision and do not necessarily invalidate the empirical phenomena, but they do prevent acceptance in the current form.

major comments (4)
  1. [V-A / Eq. (10)] The estimation robustness results in Table I do not establish that the filter tracks a genuine human belief. In Section V-A, the simulated human's ground-truth belief phi_GT is updated using 'the same kinematic moving-average formulation defined for the robot's geometric proxy (b_t^y in Eq. (10))'. The filter's geometric channel uses the same proxy. Thus the reported phi-BCE and phi-MSE numbers measure how well the filter recovers a smoothed kinematic feature generated from the same equation, not whether that feature corresponds to the human's internal belief. An independent generative model or a held-out behavioral prediction test (e.g., using hat_phi to predict real-human actions out of sample) is needed.
  2. [V-A / Eq. (16)] The in-person tracking results in Table I are also not independent evidence. Because no ground-truth internal state exists for real humans, the authors extract pseudo-labels by optimizing Eq. (16), which includes the term lambda_phi (hat_phi_t - b_t^y)^2, where b_t^y is the same geometric proxy from Eq. (10). Consequently, the pseudo-label evaluation can be satisfied by tracking the robot's own kinematic proxy under another name. This circularity should be acknowledged, and the belief channel should be validated against a measurement that does not share Eq. (10).
  3. [V-C / VI-B] BAIT-adapt's switching rule is 'success rate below 60% over the last 5 episodes OR hat_phi > 0.7'. The paper never ablates the belief condition; it only compares BAIT-adapt against BAIT-trust, BAIT-infl, and Stackelberg baselines. Therefore the reported improvements could be entirely due to the success-rate heuristic, with the belief estimate being inert or redundant. Since the title and central contribution emphasize 'Belief-Aware' behavior, an ablation is load-bearing: compare BAIT-adapt against (i) success-rate-only switching, (ii) belief-only switching, and (iii) a mismatch condition where the belief signal is corrupted. Without this, the mechanism behind the trust/task trade-off is unidentified.
  4. [V-D / VI-B] The abstract claims task performance 'comparable' to influence-only baselines, but the simulation results appear to contradict this. In Section V-D, BAIT-adapt's human lane progress (1878.0) is significantly higher than BAIT-infl (1628.7) and Belief-Entropy Stackelberg (1598.9), both p<0.001. In Section VI-B, however, BAIT-adapt is reported as equivalent to BAIT-infl (p=0.811). If human lane progress is the task-performance metric, the simulation result is inconsistent with the abstract; if a different success metric is intended, it should be explicitly defined and reported for all conditions.
minor comments (5)
  1. [III / IV-A] The transition in Eq. (1) writes z_{t+1} = f_s(s_t, a_t, z_t, phi_t), but Eq. (6) defines z_{t+1} as depending on the predicted phi_{t+1}. Please align the notation to avoid confusion about the conditioning order in the hierarchical filter.
  2. [V-D / VI-B] The terms 'task performance' and 'human lane progress' are used interchangeably. Please define the primary outcome metric explicitly and report merge success rates alongside lane progress, since the two may diverge.
  3. [VI-B] The statistical reporting is imprecise: 'across all four pairwise comparisons' should be three for three controllers, and the pairwise p-value comparing BAIT-adapt versus BAIT-trust on the trust item is not reported. Also, 'comfort median 2, discomfort median 5' appears to describe reversed-coded scales; clarify the direction.
  4. [V-D] The real-time claim '~0.11 s/step' would be more informative with a wall-clock comparison to an online MOMDP solver (e.g., POMCPOW) under the same task, rather than a standalone number.
  5. [VII] The real-world experiment is qualitative only. Please report the number of trials, the number of participants/drivers, and quantitative outcome measures (e.g., merge success, lane progress, trust ratings) to support the claim of 'consistent' yielding across repeated interactions.

Circularity Check

4 steps flagged

Filter validation is circular: simulated ground-truth belief and human pseudo-labels are built from the robot's own geometric proxy, and the simulated human is programmed to yield under the same entropy that the influence cost maximizes.

specific steps
  1. self definitional [Section V-A (Simulator and Adaptive Human Model); Eq. (10) vs. Eq. (11)]
    "Specifically, the simulated human updates its ground-truth belief φ_GT using the same kinematic moving-average formulation defined for the robot's geometric proxy (b_t^y in Eq. (10))."

    The filter's long-term belief update uses b_t^y as geometric evidence (Eq. 11). In simulation, the 'ground-truth' φ_GT is produced by the same moving-average equation (10), so Table I's φ-BCE/MSE measures how well the filter tracks a signal generated by its own evidence channel. The accuracy claim is therefore partly true by construction and does not validate that φ̂ tracks an independent human internal belief.

  2. fitted input called prediction [Section V-D (Estimation Robustness), Eq. (16); Table I In-Person User Study]
    "the third term anchors the extracted belief to the geometric yield proxy b_t^y (Eq. (10))"

    The in-person pseudo-labels used as ground truth in Table I are constructed by minimizing action NLL subject to a penalty that pulls φ̂_GT toward b_t^y, the same geometric proxy the filter uses as evidence in Eq. (11). Evaluating filter accuracy against these pseudo-labels is circular: the target is fit to the filter's own evidence channel, so low φ error in Table I is expected even if the belief estimate has no independent relation to human perception.

  3. self definitional [Section V-A (Simulator and Adaptive Human Model); Eq. (13) (influence cost)]
    "z_GT_tgt augments the baseline interpolation with two adversarial, rule-based overlays: a cautious bias that smoothly shifts the human toward a conservative prototype when belief entropy is high (H(φ_GT)>0.75), and an aggressive spike that instantly snaps to an aggressive prototype if low entropy (H(φ_GT)<0.70) persists over a moving window of 6 steps."

    The influence cost in Eq. (13) maximizes belief entropy H(φ̂) to make the human yield, based on the claim that humans 'become conservative under uncertainty.' The simulated human's z_GT is explicitly programmed to shift toward a conservative prototype when entropy is high and toward aggression when entropy is low. Hence the simulated advantage of BAIT-infl/Belief-Entropy Stackelberg over BAIT-trust in Fig. 3 (left) is produced by the same entropy→yield mapping that the controller optimizes; the simulation's task-performance comparison is self-confirming rather than an independent test of the mechanism.

  4. self citation load bearing [Section V-A (generative human model justification)]
    "Following recent findings that humans aggressively exploit predictable policies but become conservative under uncertainty [5], [6], our generative human model simulates more complex, non-linear dynamics than the robot's internal filter."

    The central behavioral premise—unpredictability/uncertainty makes humans yield—is imported from [5], [6], whose author lists overlap heavily with the present paper (Sagheb, Mun, Driggs-Campbell, Losey). This premise is then baked into the simulated human and used to motivate C_infl. The controller's simulation support therefore rests on the authors' own prior claim rather than an independently established empirical regularity, though the real-human user study provides some external support.

full rationale

The central empirical comparison (BAIT-adapt vs. BAIT-infl vs. BAIT-trust) is not entirely circular: the user study uses real participants and provides independent evidence that adaptive arbitration preserves trust relative to pure influence. However, the specifically 'belief-aware' component is circularly validated in three explicit places. First, in simulation the ground-truth long-term belief φ_GT is produced by the same kinematic moving-average equation (Eq. 10) that the robot's filter ingests as geometric evidence in Eq. (11); Table I's low φ error is therefore a check of self-consistency rather than of tracking an independent human belief. Second, for the user study, the 'ground-truth' pseudo-labels used to score the filter are fit with an explicit penalty anchoring φ to b_t^y (Eq. 10)—again the filter's own evidence channel—so the in-person estimation accuracy in Table I is constructed rather than measured. Third, the simulated human's strategy attractor is directly programmed to shift conservative when belief entropy is high and aggressive when entropy is low, which is exactly the entropy→yield mechanism C_infl optimizes (Eq. 13); the simulated task-performance advantage of influence modes is thus baked into the generative model. The premise for this generative model is itself imported from overlapping self-citations ([5], [6]). No ablation removes the belief threshold from BAIT-adapt's arbitration, so the independent contribution of belief tracking to the outcomes is untested. These are construction-level circularities in the validation of the belief channel, not merely citation continuity, but they do not erase the real-human trust results.

Axiom & Free-Parameter Ledger

11 free parameters · 5 axioms · 0 invented entities

No new physical entities are invented. The paper's contribution is algorithmic, but it introduces many hand-set parameters (11+ numeric constants and the beta-switching rule) and relies on the two-timescale generative human model as a domain assumption. The most significant ledger entry is that the simulated human's ground-truth belief uses Eq. (10), the same proxy equation the robot uses for evidence - this reduces the independence of the simulated validation of belief tracking. The pseudo-labeling objective (Eq. 16) further anchors evaluation to the robot's own priors. These are model assumptions rather than free parameters fitted to a target, so the circularity burden is moderate, not maximal.

free parameters (11)
  • BAIT-adapt arbitration switch rule = binary switching: beta=1 if success rate <60% over 5-episode window or phi_hat>0.7; else beta=0
    Hand-chosen hysteresis rule central to the BAIT-adapt result; not optimized or derived, and the paper states future work will replace it with continuous arbitration.
  • Beta kernel concentration kappa = 50.0
    Controls the drift variance of the long-term belief model; set by hand (Section V-B).
  • OU mean-reversion rate alpha_ou = 0.5
    Sets how fast the short-term strategy reverts to its attractor; hand-set.
  • OU noise scale sigma_z = 0.1
    Models human strategy stochasticity; hand-set.
  • Rationality parameter eta (Boltzmann policy) = not reported numerically in text
    In Eq. (2), controls how close human behavior is to rational; value used in experiments not stated.
  • Evidence window length W and decay rho = W=5, rho=0.7
    For the action log-likelihood window; hand-set.
  • Filter update rates and temperature = alpha_z=0.8, alpha_phi=0.25, T_phi=3.0, reset prob epsilon=0.02, reset shape c=1.0
    Filter smoothing and evidence-fusion constants; hand-set, no sensitivity analysis.
  • Entropy/directional weights w_e, w_d = w_e=0.8, w_d=0.2 (influence); w_e=0.9, w_d=0.1 (trust)
    Weights in Eqs. (13)-(14); chosen by hand to enforce w_d<<w_e.
  • CVaR risk parameter eta_r and threshold tau = eta_r=1.0, tau=2.5
    Task-safety constraint parameters; hand-set.
  • Interaction zone threshold d_interact = 25.0 m
    Gating for both evidence fusion and belief shaping; hand-set.
  • Pseudo-label weights lambda_z, lambda_phi, smoothing weight = unspecified
    Eq. (16) defines the pseudo-label objective but the relative weights are not reported, so the evaluation labels are under-specified.
axioms (5)
  • domain assumption Human behavior follows the Boltzmann policy pi_H(a|s,z) proportional to exp(-eta z^T r_bar) with z a linear strategy weight and phi affecting actions only through z.
    Eq. (2). This is the generative model used by the filter, the planner surrogate, and the paper's evaluation; if real humans deviate substantially, inference and planning degrade. The paper explicitly states BAIT is 'explicitly reliant on this structured human model'.
  • domain assumption Long-term belief phi evolves via mixture-Beta drift (Eq. 5) and short-term strategy z via OU mean-reversion toward an attractor z_tgt(phi) (Eq. 6).
    These kernels define the two-timescale hierarchy and are used both to construct the filter and to pseudo-label real data; they are not derived from independent human-behavior data.
  • domain assumption The geometric yield-belief proxy b_t^y in Eq. (10) is a valid observable surrogate for the human's true latent belief phi.
    Used as the filter's direct evidence channel and as the anchor for pseudo-labels (Eq. 16); the real-human ground-truth belief is unobservable, so all validation of phi-tracking depends on this assumed proxy.
  • domain assumption Maximum-entropy Boltzmann human behavior with decaying history (Eq. 7) and soft inertia (Eq. 8) yields a well-calibrated posterior.
    Standard approximate filtering assumptions, but the calibration claim is only tested against pseudo-labels generated from the same model family.
  • standard math Task cost is a valid proxy for merge success and safety in the CVaR constraint (Eq. 15).
    Standard stochastic MPC safety formulation; no circularity concern.

pith-pipeline@v1.3.0-alltime-deepseek · 5305 in / 5968 out tokens · 156904 ms · 2026-08-04T03:24:50.544152+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Belief-Aware Influence and Trust (BAIT): Shaping Human Belief During Repeated Human-Robot Interaction." pith.science (2026). https://pith.science/paper/ELZHV4PH

@misc{pith2026260725327,
  author       = {Pith},
  title        = {Pith review of: Belief-Aware Influence and Trust (BAIT): Shaping Human Belief During Repeated Human-Robot Interaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ELZHV4PH}},
  note         = {Machine review of arXiv:2607.25327}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Repeated human-robot interaction (HRI) requires proactively accounting for humans who continually adapt to evolving beliefs about the robot. Prior frameworks often treat encounters as isolated events, suffering cumulative task performance decay as human perception drifts, or maintain long-term influence through erratic, unpredictable behavior that erodes perceived human trust and relies on computationally unscalable formulations. To address these gaps, we introduce the Belief- Aware Influence and Trust (BAIT) controller. BAIT integrates a hierarchical particle filter, which infers both fast human strategic shifts and slow perceptual belief updates, with a belief-aware Model Predictive Path Integral planner. BAIT explicitly optimizes the trade-off between long-horizon influence and human trust, while enforcing immediate task performance as a strict constraint. Across simulations, a human-subject study, and a real-world GEM vehicle deployments in repeated lane-merging scenarios, BAIT achieves task performance comparable to baselines that optimize long-term influence through unpredictability while yielding significantly higher user trust. The video demonstrating our experiments is available at https://youtu.be/9o4GqKLWDCw.

Figures

Figures reproduced from arXiv: 2607.25327 by Dylan P. Losey, Katherine Driggs-Campbell, Mahsa Golchoubian, Shahabedin Sagheb, Tianhao Ji, Yan Bai, Ye-Ji Mun.

Figure 1
Figure 1. Figure 1: Repeated lane-merging encounters deployed on Polaris GEM vehicles. Our BAIT-adapt robot (white) explicitly navigates the internal trade￾off between influence to assert its lane right-of-way and preserving human trust. As shown in the corresponding trajectory rollouts (right), the BAIT controller (solid red) successfully executes the merge across repeated 10 episodes while consistently prompting the human d… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our BAIT controller. (a) Our BAIT controller has two tightly coupled modules: a hierarchical particle filter that tracks the joint posterior over (ϕ, z) from observed interaction history, and an MPPI planner that rolls this posterior forward through a deterministic belief-propagation using the posterior mean (ϕ, ˆ zˆ). This planner optimizes an explicit influence-trust trade-off while enforcing… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of Human Lane Progress. Overall lane progress for the BAIT variants (BAIT-trust, BAIT-infl, and BAIT-adapt) against Stackelberg baselines (Stackelberg, Belief-Entropy Stackelberg) when interacting with a belief-aware simulated human (left), alongside the overall (center) and episode￾by-episode (right) progress of in-person users interacting with the three BAIT variants. Both the simulated and in… view at source ↗
Figure 4
Figure 4. Figure 4: Evolution of Interaction trajectories. Comparison of robot (solid red) and simulated human (dashed blue) trajectories during early (Eps 1 − 5) and late (Eps 16 − 20) interactions. Results are aggregated over 30 independent trials across five control modes. While standard Stackelberg and BAIT-trust initially secure the lane merge, they fail to prevent the human from becoming aggressive over repeated interac… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 2 linked inside Pith

  1. [1]

    Occlusion-aware crowd navigation using people as sensors,

    Y .-J. Mun, M. Itkina, S. Liu, and K. Driggs-Campbell, “Occlusion-aware crowd navigation using people as sensors,” inProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 12 031–12 037

    2023

  2. [2]

    Learning latent representations to influence multi-agent interaction,

    A. Xie, D. Losey, R. Tolsma, C. Finn, and D. Sadigh, “Learning latent representations to influence multi-agent interaction,” inConference on Robot Learning (CoRL), 2021, pp. 575–588

    2021

  3. [3]

    Mixed strategy nash equilibrium for crowd navigation,

    M. Muchen Sun, F. Baldini, K. Hughes, P. Trautman, and T. Murphey, “Mixed strategy nash equilibrium for crowd navigation,”The Interna- tional Journal of Robotics Research (IJRR), vol. 44, no. 7, pp. 1156– 1185, 2025

    2025

  4. [4]

    A survey on trajectory-prediction methods for autonomous driving,

    Y . Huang, J. Du, Z. Yang, Z. Zhou, L. Zhang, and H. Chen, “A survey on trajectory-prediction methods for autonomous driving,”IEEE Transactions on Intelligent Vehicles, vol. 7, no. 3, pp. 652–674, 2022

    2022

  5. [5]

    A unified framework for robots that influence humans over long-term interaction,

    S. Sagheb, S. Parekh, R. Pandya, Y .-J. Mun, K. Driggs-Campbell, A. Ba- jcsy, and D. P. Losey, “A unified framework for robots that influence humans over long-term interaction,”arXiv preprint arXiv:2503.14633, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Towards robots that influence humans over long-term interaction,

    S. Sagheb, Y .-J. Mun, N. Ahmadian, B. A. Christie, A. Bajcsy, K. Driggs-Campbell, and D. P. Losey, “Towards robots that influence humans over long-term interaction,” inProceedings of the IEEE In- ternational Conference on Robotics and Automation (ICRA), 2023, pp. 7490–7496

    2023

  7. [7]

    On legible and pre- dictable robot navigation in multi-agent environments,

    J.-L. Bastarache, C. Nielsen, and S. L. Smith, “On legible and pre- dictable robot navigation in multi-agent environments,” inProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 5508–5514

    2023

  8. [8]

    Legibility and predictabil- ity of robot motion,

    A. D. Dragan, K. C. Lee, and S. S. Srinivasa, “Legibility and predictabil- ity of robot motion,” inProceedings of the ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2013, pp. 301–308

    2013

  9. [9]

    Planning for autonomous cars that leverage effects on human actions,

    D. Sadigh, S. Sastry, S. A. Seshia, and A. D. Dragan, “Planning for autonomous cars that leverage effects on human actions,” inRobotics: Science and Systems (RSS), 2016

    2016

  10. [10]

    Hierarchical game-theoretic planning for autonomous vehicles,

    J. F. Fisac, E. Bronstein, E. Stefansson, D. Sadigh, S. S. Sastry, and A. D. Dragan, “Hierarchical game-theoretic planning for autonomous vehicles,” inProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2019, pp. 9590–9596

    2019

  11. [11]

    Social behavior for autonomous vehicles,

    W. Schwarting, A. Pierson, J. Alonso-Mora, S. Karaman, and D. Rus, “Social behavior for autonomous vehicles,”Proceedings of the National Academy of Sciences, vol. 116, no. 50, pp. 24 972–24 978, 2019

    2019

  12. [12]

    Learning latent representations to co-adapt to humans,

    S. Parekh and D. P. Losey, “Learning latent representations to co-adapt to humans,”Autonomous Robots, vol. 47, no. 6, pp. 771–796, 2023

    2023

  13. [13]

    Stackelberg punishment and bully-proofing autonomous vehicles,

    M. Cooper, J. K. Lee, J. Beck, J. D. Fishman, M. Gillett, Z. Papakipos, A. Zhang, J. Ramos, J. A. Shah, and M. L. Littman, “Stackelberg punishment and bully-proofing autonomous vehicles,” inInternational Conference on Social Robotics (ICSR), 2019, pp. 368–377

    2019

  14. [14]

    Continual learning through human-robot interaction: Human perceptions of a continual learning robot in repeated interactions,

    A. Ayub, Z. De Francesco, P. Holthaus, C. L. Nehaniv, and K. Daut- enhahn, “Continual learning through human-robot interaction: Human perceptions of a continual learning robot in repeated interactions,” International Journal of Social Robotics, vol. 17, no. 2, pp. 277–296, 2025

    2025

  15. [15]

    Trust-aware em- bodied bayesian persuasion for mixed-autonomy,

    S. Peng, K. Driggs-Campbell, and R. Dong, “Trust-aware em- bodied bayesian persuasion for mixed-autonomy,”arXiv preprint arXiv:2509.15404, 2025

    arXiv 2025

  16. [16]

    Learning latent representations to influence multi-agent interaction,

    A. Xie, D. P. Losey, R. Tolsma, C. Finn, and D. Sadigh, “Learning latent representations to influence multi-agent interaction,” inConference on Robot Learning (CoRL), 2021, pp. 575–588

    2021

  17. [17]

    Learning vision- based pursuit-evasion robot policies,

    A. Bajcsy, A. Loquercio, A. Kumar, and J. Malik, “Learning vision- based pursuit-evasion robot policies,” inProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 9197–9204

    2024

  18. [18]

    Hierar- chical intention tracking with switching trees for real-time adaptation to dynamic human intentions during collaboration,

    Z. Huang, Y .-J. Mun, F. C. Pouria, and K. Driggs-Campbell, “Hierar- chical intention tracking with switching trees for real-time adaptation to dynamic human intentions during collaboration,”arXiv preprint arXiv:2506.07004, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    A perspective on incen- tive design: Challenges and opportunities,

    L. J. Ratliff, R. Dong, S. Sekar, and T. Fiez, “A perspective on incen- tive design: Challenges and opportunities,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 2, no. 1, pp. 1–34, 2018. 9

    2018

  20. [20]

    Learning with opponent-learning awareness,

    J. Foerster, R. Y . Chen, M. Al-Shedivat, S. Whiteson, P. Abbeel, and I. Mordatch, “Learning with opponent-learning awareness,” inInter- national Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2018, pp. 122–130

    2018

  21. [21]

    Reinforcement learning based control of imitative policies for near-accident driving,

    Z. Cao, E. Biyik, W. Z. Wang, A. Raventos, A. Gaidon, G. Rosman, and D. Sadigh, “Reinforcement learning based control of imitative policies for near-accident driving,” inRobotics: Science and Systems (RSS), 2020

    2020