Environment Design for Reliable Shared Autonomy with Probabilistic Guarantees

Alessandro Roncone; Bradley Hayes; Gyanig Kumar; Heyang Huang; Himanshu Gupta; Yi-Shiuan Tung

REVIEW 4 major objections 4 minor 18 references

Environment Design for Reliable Shared Autonomy with Probabilistic Guarantees

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that arranging objects in a workspace so that a margin condition holds guarantees, with probability at least 1−α, that the shared-autonomy robot infers the user's intended goal.

desk verdict The workspace-design idea is worth taking seriously, but the paper's headline guarantee is not established — the theorem assumes its own conclusion, and the Hard-B numbers contradict the text. read the letter →

arxiv 2607.15487 v1 pith:AKFHFLMW submitted 2026-07-16 cs.RO

Yi-Shiuan Tung , Himanshu Gupta , Gyanig Kumar , Heyang Huang , Bradley Hayes , Alessandro Roncone This is my paper

classification cs.RO

keywords sharedautonomyworkspacedesignintentinferenceprobabilisticguaranteegoaldisambiguationquality-diversityoptimizationtrajectorymarginslackhuman-robotinteraction

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

The paper argues that the physical arrangement of objects on a table is not a fixed background but a design variable that determines how easily a shared-autonomy robot can tell which object the human user is reaching for. It defines a margin for each pair of candidate goals that combines the nominal log-posterior separation under a rational-choice observer with a penalty for how much execution noise can blur that separation, and proves that if this 'trajectory margin slack' is positive for every alternative goal, the correct goal is identified with probability at least 1−α. The authors optimize this worst-case slack over object positions and orientations using a quality-diversity evolutionary search, and in simulation the optimized layouts outperform random baselines on accuracy across six tabletop scenarios, including a jump from 25% to 54% in the most cluttered case. The result matters because it makes workspace layout a complementary axis for reliability in shared autonomy, alongside better inference algorithms.

What carries the argument

The load-bearing object is the trajectory margin slack s_g(θ) = m_g(θ) − Φ^{−1}(1−α_g)√v_g(θ), a scalar for each alternative goal that balances nominal goal separability against noise-induced variance. A positive slack for all alternatives is shown to be sufficient for the 1−α correctness guarantee; the workspace optimization then maximizes the worst-case slack across all goals using a quality-diversity search over (x, y, yaw) placements, with behavior descriptors capturing spatial spread and centroid offset.

What would settle it

Run the optimized layouts in a physics simulator under the actual 20 Hz velocity controller, measure the empirical argmax accuracy over many noise draws, and test whether it falls below the 1−α bound for layouts with positive slack; also directly estimate the terminal-state covariance and check whether it matches the isotropic Gaussian assumption. A significant violation of either would falsify the guarantee's applicability.

Watch

Extended reading notes

Core claim

The paper's central discovery is a sufficient condition connecting geometry to correctness: for a fixed true goal, if for every alternative goal the trajectory margin slack s_g = m_g − Φ^{−1}(1−α_g)√v_g is positive, then the probability that the observer picks the true goal is at least 1−α. Here m_g is the nominal log-posterior margin under the rational-choice goal model, and v_g is the variance of that margin induced by linearizing the cost difference around the nominal terminal state under Gaussian control noise. The paper then shows that maximizing the worst-case slack over object placements yields near-perfect argmax accuracy in simulation (100% in most scenarios, 97% in the hardest eigh

Load-bearing premise

The guarantee requires that the end-effector's terminal state be Gaussian with covariance Tσ_u^2Δt^2I and that the linearization error of the cost difference be negligible inside the (1−α) noise ball; the experiments use a Jacobian-pseudoinverse controller that does not match this model, and the linearization condition is never evaluated.

Editorial extensions

If this is right

If the slack condition holds in a real deployment, a layout can be certified at design time: the robot will infer the user's intended goal with at least 1−α probability, independent of the particular inference update once the cost model is fixed.
The approach directly reduces ambiguity in cluttered scenes: in the eight-object scenario, optimized layouts raised argmax accuracy from 25% to 54% relative to random layouts.
Optimized layouts also disambiguate faster in moderate settings (e.g., time-to-inference dropped from about 6.9 s to 4.1 s in one three-object scene), meaning assistance kicks in sooner.
Because the search returns a diverse archive of high-quality layouts, designers can choose among certified layouts to satisfy additional preferences (aesthetics, reachability, task semantics) without losing the probabilistic guarantee.
The formulation extends naturally to multi-stage real-world tasks by optimizing only the goal subset relevant to the current stage, as demonstrated in tea-making, block sorting, and assistive-feeding prototypes.

Reading between the lines

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

The 1−α guarantee is derived under an idealized Gaussian integrator noise model; under the actual Jacobian-pseudoinverse velocity controller used in the experiments the terminal-state distribution was not verified, so the bound should be treated as a design-time heuristic until the covariance and linearization-error conditions are empirically checked.
The optimization assumes users follow a straight-line policy toward the goal with isotropic Gaussian noise of known scale; if human teleoperation deviates from this policy (e.g., curved paths or speed-dependent noise), the calculated slack may not reflect true separability.
A direct testable extension: in a user study, measure empirical argmax accuracy on optimized layouts and compare to the 1−α bound; if the bound holds, the certificate becomes a practically useful specification rather than a theoretical one.
The margin-slack objective could be coupled with robot base placement or joint-space redundancy, extending environment design from object arrangement to the full shared-autonomy configuration.

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

Colleague — quick read of arXiv:2607.15487. The core idea is genuinely sensible: instead of only improving the inference algorithm, rearrange the tabletop so that candidate goals are more separable under noisy human input. That extends the authors' own deterministic legibility work to the noisy setting, and the margin-based chance-constraint formulation (maximize worst-case slack) is a reasonable way to pose it. The MAP-Elites/CMA-ME optimization is appropriately matched to the non-convex layout space, and the simulation setup is described with enough detail to be reproducible in principle.

The problem is the central theorem. Theorem 1 claims that positive slack s_g(θ) for all alternatives implies correct goal inference with probability ≥ 1−α. The proof goes through only if each pairwise margin M_g is exactly Gaussian. Assumption 2 just asserts this; it does not follow from the linearization in Eq. (8). Appendix A bounds the linearization error on a high-probability ball (|R_g| ≤ L_g r_α²/2), but a bounded pointwise error does not make M_g Gaussian and does not bound the tail probability P(M_g < 0) that actually appears in the union bound. A non-Gaussian distribution with the same mean and variance can have a much heavier lower tail. So the theorem effectively assumes the conclusion's key premise. That is a load-bearing gap, not a cosmetic one.

Second, the covariance in Eq. (7) is derived for a first-order integrator with i.i.d. Gaussian noise. The actual simulator is a 20 Hz Jacobian pseudoinverse velocity controller with speed saturation; the terminal state is not Gaussian. The linearization-error condition (Eq. 17) is never evaluated anywhere in the experiments, so we have no evidence the approximation holds in the regimes tested.

Third, the results section is internally inconsistent. The text says optimized layouts maintain 'near-perfect performance (100% in most cases and 97% in Hard-B),' but Table II reports Hard-B ME-optimized as 54% accuracy with negative slack (−0.999). Negative slack means the sufficient condition of Theorem 1 is not met, so the guarantee does not apply; the 54% number is actually consistent with that, but the prose claim is not. That is the one scenario where the guarantee would matter most, and the paper's own data undercuts it.

The real-world section is a qualitative demo with no numbers, so it does not add evidentiary weight.

Where credit is due: the paper is well-structured, the related work is honest about what is new, the parameters are reported, and the empirical trend (optimization helps accuracy in simulation) is plausible even if the guarantee is unproven.

Bottom line: this is a conference-worthy idea in need of a major repair. A serious referee should not reject it out of hand — they should send it back for the authors to either fix the Gaussian/tail argument, validate Eq. (17) in the simulator, align the theory with the actual controller, and reconcile the text with the table. If the theory can be repaired or the claims appropriately weakened, it could be a solid contribution.

I would bring it to a reading group to talk through the theorem gap as a case study in overclaiming. I would not cite it in its current form.

Referee Report

4 major / 4 minor

Summary. The paper proposes to optimize the physical layout of objects in a shared-autonomy workspace to improve goal-inference reliability. The authors model user input as a noisy unit vector toward the goal, use a Boltzmann-rational observer, and define a margin slack s_g(θ). Theorem 1 claims that if the slack is positive for all alternatives, the probability of correct goal inference is at least 1−α. Layouts are optimized with CMA-ME in SE(2) (x,y,yaw). Experiments in MuJoCo across six scenarios compare random layouts with optimized layouts, reporting argmax accuracy and time-to-inference. A qualitative real-world demonstration is also presented. The central issue is that the guarantee rests on unverified Gaussianity and linearization assumptions, and the experimental text contradicts the table.

Significance. The idea of treating workspace design as a decision variable for inference is appealing and complements prior algorithmic work on intent inference. The use of MAP-Elites to produce diverse high-slack layouts and the inclusion of a real-world tabletop demonstration are strengths. If the guarantee were rigorously established and validated, the paper would offer a useful design-time certificate. However, the main theoretical contribution is not established: Theorem 1 assumes the margin is Gaussian, which is essentially the property to be proved, and the experiments neither check the linearization-error condition nor use a controller consistent with the covariance model. The internal contradiction between the prose and Table II further undermines confidence. The paper is therefore not ready for publication in its present form.

major comments (4)

[§II-B-d, Appendix A, Eq. (17)] Theorem 1 is not derived from the noise model; it assumes the key conclusion. Assumption 2 states that M_g is Gaussian, but Assumption 1 (Gaussian x_T) plus the first-order expansion in Eq. (8) does not imply Gaussianity of M_g. Appendix A only bounds the remainder |R_g| ≤ L_g r_α²/2 on a (1−α)-probability set. A bounded pointwise error does not make the tail Gaussian: a non-Gaussian variable with the same mean and variance can have P(M_g<0) much larger than Φ(−m_g/√v_g). Moreover, Eq. (17) ('L_g r_α²/2 ≪ m_g') is a heuristic order-of-magnitude condition, not a quantitative sufficient condition for the Gaussian tail bound, and it is never checked in the experiments. Thus the stated probabilistic guarantee is conditional on an assumption that is neither proven nor validated.
[§III-A-d, Eq. (7), Eq. (8)] The covariance in Eq. (7), Σ_xT = T σ_u² Δt² I, is derived for a first-order integrator with i.i.d. Gaussian control noise. The simulation instead uses a 20 Hz Jacobian pseudoinverse velocity controller with speed saturation and IK-based execution (Section III-A-d). Nonlinear feedback, clipping, and the projection through the pseudoinverse make the terminal state non-Gaussian, and no validation of the Gaussian assumption is provided. In addition, the cost C(ξ,g) in Eq. (2) depends on the full trajectory, but the linearization in Eq. (8) is only with respect to x_T; the distribution of the path is not captured by the terminal-state covariance. The paper therefore does not establish that the conditions of Theorem 1 hold in the experimental system.
[§II-A vs §III-A-e] The theoretical margin is defined for the cost C(ξ,g) = trajectory length + remaining distance-to-go (Eq. (2)), but the simulated observer uses a different 'path-efficiency cost ratio using an SE(3) distance metric with rotation weight λ_R=0.04' (Section III-A-e). Since the slack and the guarantee are derived for the former cost, the optimized layouts are not guaranteed to maximize the actual observer's margin. This mismatch is not discussed and further severs the link between Theorem 1 and the experimental results.
[§IV-A-a, Table II] The reported results are internally contradictory. The text states that random layouts degrade from 88% (Easy) to 67% (Hard-B) and that optimized layouts maintain near-perfect performance, '100% in most cases and 97% in Hard-B.' Table II reports random Hard-B argmax accuracy of 25% and ME-optimized Hard-B of 54%, with slack −0.999. The negative slack means the sufficient condition of Theorem 1 is not satisfied, so no ≥95% guarantee can be claimed for Hard-B. The prose and the table cannot both be correct; as printed, the table undermines the paper's central experimental claim.

minor comments (4)

[Fig. 2 caption and Appendix B-a] The caption calls the layouts 'SE(3)-optimized', but the optimization is over (x,y,yaw) only. Please reword to planar/SE(2) optimization to avoid confusion.
[Table II] The ME-optimized row appears to be a single elite; report the number of runs and, where relevant, variability. Several random rows include mean±std but the optimized rows do not.
[§IV-C] The real-world section is labeled a 'demonstration' but does not report quantitative results; this is fine, but the text should not imply a user study, which is deferred to future work.
[Miscellaneous] Typographical errors: 'guaranties' in Section II-B-e, 'Quality-Diversty' in Appendix B title, and 'valided' in Section III-A-f.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 1 is a conditional implication from explicit assumptions, and the empirical evaluations do not reduce to the optimization objective by construction.

full rationale

I examined Theorem 1 (Eq. 10), the slack definition (Eq. 9), the workspace design objective (Eq. 11), the MAP-Elites optimization, and the experimental protocol. No load-bearing step reduces to its own input by construction. Theorem 1 is stated as a conditional: under Assumptions 1–3, positive slack s_g(θ)>0 implies P(ĝ=g*)≥1−α via Gaussian tail bounds and a union bound. The Gaussianity of the margin M_g is explicitly listed as Assumption 2 rather than derived from the conclusion, so the theorem is a valid implication, not a circular derivation. Appendix A only bounds the linearization remainder; whether that bound suffices to justify Assumption 2 is a correctness/validation question, not a circularity. The simulations use the same Boltzmann-rational observer and Gaussian noise model as the analytic derivation, making them self-consistency checks, but the reported argmax accuracies are not computed from the slack values, and the one case with negative slack (Hard-B ME-optimized, slack −0.999) shows degraded accuracy (54%), which is qualitatively consistent with the sufficiency claim rather than forced by it. There is a reporting discrepancy (text says 97% in Hard-B while Table II says 54%), but that is a data/consistency issue, not circularity. The only self-citations [6],[7] are prior workspace-optimization papers cited as related work and are not load-bearing for Theorem 1 or the experimental results. No circular steps were found.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a Boltzmann-rational observer with a hand-picked rationality coefficient, a Gaussian noise model with a hand-picked variance, and an unverified linearization-error assumption. The strongest load-bearing premise is the Gaussian terminal-state covariance, which does not match the actual controller used in the experiments.

free parameters (4)

rationality coefficient β = 5
Controls the sharpness of the Boltzmann-rational observer (Eq. 2). Chosen by hand; directly affects margin magnitude and the slack condition.
joystick noise standard deviation σ_v = 0.03 m/s
Assumed noise level in the simulated user input (Eq. 1). Not measured from data; determines the terminal-state covariance and hence the guarantee.
rotation weight λ_R = 0.04
Weight in the SE(3) distance metric used by the observer in simulation; chosen by hand.
commit threshold p_thresh = 0.9
Posterior probability at which the system commits to a goal; arbitrary design choice that affects timing and accuracy.

assumptions (4)

domain assumption Human input follows a bounded Gaussian noise model: u_t = π_h(x_t, g*) + ε_t, ε_t ~ N(0, Σ) (Eq. 1).
All subsequent derivations and the guarantee depend on this noise model. Not validated against real user data.
domain assumption The robot's belief over goals follows the Boltzmann-rational model with cost C(ξ, g) = trajectory length + remaining distance (Eq. 2).
The inference framework is assumed to be the observer; the guarantee is defined relative to this observer.
domain assumption The terminal state is Gaussian with covariance T σ_u^2 Δt^2 I (Eq. 7).
This holds only for a first-order integrator with i.i.d. Gaussian control noise and a fixed nominal policy; the actual controller is a Jacobian pseudoinverse velocity controller, so this is not satisfied.
ad hoc to paper The linearization error R_g of δc is negligible within a (1−α)-probability neighborhood, satisfying L_g r_α^2/2 ≪ m_g (Eq. 17).
The paper does not verify this condition in simulation. It is introduced solely to justify the Gaussian margin approximation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Environment Design for Reliable Shared Autonomy with Probabilistic Guarantees." pith.science (2026). https://pith.science/paper/AKFHFLMW

@misc{pith2026260715487,
  author       = {Pith},
  title        = {Pith review of: Environment Design for Reliable Shared Autonomy with Probabilistic Guarantees},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKFHFLMW}},
  note         = {Machine review of arXiv:2607.15487}
}

read the original abstract

Shared autonomy enables humans and robots to collaboratively perform tasks by combining human input with autonomous assistance. Most prior work focuses on improving intent inference under a fixed environment, overlooking how workspace design itself affects inference difficulty. We observe that the physical arrangement of objects directly influences the separability of candidate goals under noisy user inputs. We formulate workspace design as an optimization problem and derive a probabilistic correctness guarantee under a bounded noise model. Through simulation experiments across multiple tabletop scenarios, we show that optimized layouts improve goal inference reliability and reduce ambiguity compared to baseline arrangements. We further demonstrate a real-world shared autonomy system that integrates the proposed inference framework. This highlights the role of environment design as a complementary axis for improving shared autonomy systems.

Figures

Figures reproduced from arXiv: 2607.15487 by the authors.

**Figure 2.** Random vs. SE(3)-optimized workspace layouts. Each pair shows a random baseline (left) and an SE(3)-optimized layout (right) produced by MAP-Elites over (x,y,yaw) using the trajectory-margin objective. All layouts satisfy identical geometric constraints. By optimizing in SE(3), the method jointly places and orients objects to maintain separability of reachable grasp trajectories across M candidate goals. Tasks share… view at source ↗

**Figure 3.** Illustration of three shared autonomy scenarios: (a) grabbing snacks for tea, (b) sorting LEGO blocks into containers, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 3 linked inside Pith

[1]

Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction , pages=

Workspace optimization techniques to improve prediction of human motion during human-robot collaboration , author=. Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction , pages=

2024
[2]

2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI) , pages=

Legibility and predictability of robot motion , author=. 2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI) , pages=. 2013 , organization=

2013
[3]

Proceedings of the 2020 genetic and evolutionary computation conference , pages=

Covariance matrix adaptation for the rapid illumination of behavior space , author=. Proceedings of the 2020 genetic and evolutionary computation conference , pages=

2020
[5]

Journal of global optimization , volume=

Differential evolution--a simple and efficient heuristic for global optimization over continuous spaces , author=. Journal of global optimization , volume=. 1997 , publisher=

1997
[7]

IEEE Access , year=

Human following and guidance by autonomous mobile robots: A comprehensive review , author=. IEEE Access , year=
[8]

Surgical Clinics , volume=

The da Vinci telerobotic surgical system: the virtual operative field and telepresence surgery , author=. Surgical Clinics , volume=. 2003 , publisher=

2003
[10]

The International Journal of Robotics Research , volume=

Shared autonomy via hindsight optimization for teleoperation and teaming , author=. The International Journal of Robotics Research , volume=. 2018 , publisher=

2018
[11]

and Roncone, Alessandro and Hayes, Bradley , title =

Tung, Yi-Shiuan and Luebbers, Matthew B. and Roncone, Alessandro and Hayes, Bradley , title =. 6th International Workshop on Virtual, Augmented, and Mixed-Reality for Human-Robot Interactions (VAM-HRI) , year =

Show all 18 references

[12]

Eirale, M

A. Eirale, M. Martini, and M. Chiaberge, ``Human following and guidance by autonomous mobile robots: A comprehensive review,'' IEEE Access, 2025

2025
[13]

Z. Fu, P. Song, Y. Hu, and R. Detry, ``Tasc: Task-aware shared control for teleoperated manipulation,'' arXiv preprint arXiv:2509.10416, 2025

2025 arXiv
[14]

G. H. Ballantyne and F. Moll, ``The da vinci telerobotic surgical system: the virtual operative field and telepresence surgery,'' Surgical Clinics, vol. 83, no. 6, pp. 1293--1304, 2003

2003
[15]

H. Liu, R. Shah, S. Liu, J. Pittenger, M. Seo, Y. Cui, Y. Bisk, R. Mart \' n-Mart \' n, and Y. Zhu, ``Casper: Inferring diverse intents for assistive teleoperation with vision language models,'' arXiv preprint arXiv:2506.14727, 2025

2025 arXiv
[16]

Javdani, H

S. Javdani, H. Admoni, S. Pellegrinelli, S. S. Srinivasa, and J. A. Bagnell, ``Shared autonomy via hindsight optimization for teleoperation and teaming,'' The International Journal of Robotics Research, vol. 37, no. 7, pp. 717--742, 2018

2018
[17]

Y.-S. Tung, M. B. Luebbers, A. Roncone, and B. Hayes, ``Workspace optimization techniques to improve prediction of human motion during human-robot collaboration,'' in Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, 2024, pp. 743--751

2024
[18]

1em plus 0.5em minus 0.4em ACM, Mar 2023

------, ``Improving human legibility in collaborative robot tasks through augmented reality and workspace preparation,'' in 6th International Workshop on Virtual, Augmented, and Mixed-Reality for Human-Robot Interactions (VAM-HRI). 1em plus 0.5em minus 0.4em ACM, Mar 2023

2023
[19]

A. D. Dragan, K. C. Lee, and S. S. Srinivasa, ``Legibility and predictability of robot motion,'' in 2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI). 1em plus 0.5em minus 0.4em IEEE, 2013, pp. 301--308

2013
[20]

Mouret and J

J.-B. Mouret and J. Clune, ``Illuminating search spaces by mapping elites,'' arXiv preprint arXiv:1504.04909, 2015

2015 arXiv
[21]

M. C. Fontaine, J. Togelius, S. Nikolaidis, and A. K. Hoover, ``Covariance matrix adaptation for the rapid illumination of behavior space,'' in Proceedings of the 2020 genetic and evolutionary computation conference, 2020, pp. 94--102

2020

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction , pages=

Workspace optimization techniques to improve prediction of human motion during human-robot collaboration , author=. Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction , pages=

2024

[2] [2]

2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI) , pages=

Legibility and predictability of robot motion , author=. 2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI) , pages=. 2013 , organization=

2013

[3] [3]

Proceedings of the 2020 genetic and evolutionary computation conference , pages=

Covariance matrix adaptation for the rapid illumination of behavior space , author=. Proceedings of the 2020 genetic and evolutionary computation conference , pages=

2020

[4] [5]

Journal of global optimization , volume=

Differential evolution--a simple and efficient heuristic for global optimization over continuous spaces , author=. Journal of global optimization , volume=. 1997 , publisher=

1997

[5] [7]

IEEE Access , year=

Human following and guidance by autonomous mobile robots: A comprehensive review , author=. IEEE Access , year=

[6] [8]

Surgical Clinics , volume=

The da Vinci telerobotic surgical system: the virtual operative field and telepresence surgery , author=. Surgical Clinics , volume=. 2003 , publisher=

2003

[7] [10]

The International Journal of Robotics Research , volume=

Shared autonomy via hindsight optimization for teleoperation and teaming , author=. The International Journal of Robotics Research , volume=. 2018 , publisher=

2018

[8] [11]

and Roncone, Alessandro and Hayes, Bradley , title =

Tung, Yi-Shiuan and Luebbers, Matthew B. and Roncone, Alessandro and Hayes, Bradley , title =. 6th International Workshop on Virtual, Augmented, and Mixed-Reality for Human-Robot Interactions (VAM-HRI) , year =