Pith. sign in

REVIEW 4 major objections 4 minor 35 references

SIL: Symbiotic Interactive Learning for Language-Conditioned Human-Agent Co-Adaptation

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing one-way command execution with bidirectional co-adaptation in a shared latent belief space raises task completion to 90.4% and belief alignment to approximately 0.83, about 20 points above the best ablation.

desk verdict The belief alignment metric compares two embeddings from the same encoder, so the headline co-adaptation claim is under-evidenced; still, the engineering framework is coherent and deserves referee time. read the letter →

arxiv 2511.05203 v3 pith:KSWOJCYW submitted 2025-11-07 cs.RO

Linus Nwankwo , Bjoern Ellensohn , Christian Rauch , Elmar Rueckert This is my paper
classification cs.RO
keywords human-robotinteractionco-adaptationsharedlatentspacebeliefalignmentcontinuallearningclarificationdialoguepreferencefoundationmodels
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 current language-conditioned human-robot interaction frameworks are one-way: the human specifies, the robot executes. It introduces Symbiotic Interactive Learning (SIL), in which both parties maintain evolving belief states in a shared latent task space, with the agent explicitly modeling the human's belief and updating its own accordingly. The central claim is that this mutual belief update, plus memory and anti-forgetting safeguards, enables proactive clarification, preference learning, and stable long-term interaction. In simulated and real-world embodied tasks, SIL reports 90.4% mean task completion and a belief alignment of about 0.83, roughly 20 percentage points above its strongest ablation. A sympathetic reader would take this as evidence that foundation models can be grounded for human-robot interaction through co-adaptation rather than passive command following.

What carries the argument

The central object is the shared latent task space Z in R^d, with belief states for human and agent containing latent embeddings, confidence scalars, uncertainty representations, memory, and (for the human) a preference model. The engine is the bidirectional influence update: tanh-transformed cross-embeddings produce influence vectors, and each participant's embedding is updated as a weighted mix of its own state, new observations, and the other's influence, scaled by the interaction success signal. The alignment score rho = (1 + cos(z_H, z_A)) / 2 times c_H * c_A triggers clarification when low. Supporting mechanisms are a triplet-loss latent encoder, episodic and semantic memories with bel

What would settle it

A falsifying observation would be a controlled user study in which human partners interact with the system and their self-reported or behaviorally measured understanding (e.g., ratings of the agent's interpretations, ability to predict its next action, or spontaneous rephrasing following clarifications) fails to correlate with the alignment score; if the score rises while human understanding does not, the metric is not measuring what it claims.

Watch

Extended reading notes

Core claim

In SIL, the agent maintains two latent belief states: its own and an estimate of the human's. Both evolve through coupled update equations: influence vectors derived from the other's embedding modulate each update, with a success signal strengthening agent adaptation on success and human adaptation on failure. A confidence-weighted cosine alignment score monitors the agreement between the two belief states; when it falls below a threshold, the agent initiates clarification. The paper claims this mechanism, trained with a triplet-loss encoder and protected by elastic weight consolidation, produces stable convergence to a belief alignment of about 0.83 and a 90.4% task completion rate, with th

Load-bearing premise

The central claim rests on the assumption that the agent's internally estimated human belief state faithfully reflects the actual human's belief; if it does not, the measured alignment of about 0.83 is self-consistency, not genuine co-adaptation.

Editorial extensions

If this is right

  • If valid, autonomous agents can move beyond one-shot command execution to proactive clarification and suggestion, reducing the human's cognitive load in ambiguous situations.
  • The framework provides a quantitative, continuous measure of human-agent alignment that could serve as a real-time indicator of interaction quality and a trigger for intervention.
  • The memory and continual-learning components imply that robots can accumulate user preferences and task knowledge across sessions without overwriting earlier learning, enabling genuine long-term personalization.
  • The roughly 20-point improvement over the strongest ablation attributes most of the gain to the co-adaptation loop, suggesting that bidirectional belief modeling, not just better language grounding, is what drives performance.
  • The method is task-domain general—covering instruction following, information retrieval, reasoning, dialogue, and preference learning—so it points to a common substrate for co-adaptive human-robot interaction rather than bespoke per-task designs.

Reading between the lines

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

  • The alignment score compares the agent's estimate of the human's belief with the agent's own belief; without an independent measure of the human's actual belief, the reported 0.83 may reflect internal consistency of one model rather than true human-agent convergence. A natural next step is a user study with human-reported understanding or behavioral adaptation metrics.
  • The belief update equations are manually specified with fixed mixing coefficients; a testable extension is to learn these dynamics end-to-end from interaction data, which could reveal whether the hand-designed update is optimal or merely sufficient.
  • The framework's 'human adaptation' side is simulated by the agent's model of the human; treating the human as a real adaptive partner—measured via rephrasing, correction rates, or subjective workload—is an open empirical question.
  • If co-adaptation proves robust, the same shared latent-space mechanism could extend beyond robot teleoperation to multi-agent teams, where each agent maintains a belief over the others' beliefs to coordinate collaborative tasks.
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.

Referee Report

4 major / 4 minor

Summary. The paper proposes Symbiotic Interactive Learning (SIL), a framework in which a human and an autonomous agent co-adapt within a shared latent task space. The agent maintains an explicit belief state over the task and an estimate of the human's belief state, updated through a bidirectional influence mechanism. The framework uses a triplet-trained latent encoder, episodic and semantic memory, and elastic weight consolidation for continual learning. Experiments are reported on simulated and real-world embodied tasks in five categories (instruction following, memory-based retrieval, query reasoning, dialogue, and preference learning), with a claimed task completion rate of 90.4% and a belief alignment score of ρ≈0.83, about 20 points above ablations.

Significance. If the empirical claims were supported, SIL would represent a meaningful step toward bidirectional human-agent interaction, addressing a recognized limitation of current one-way command-and-control HRI systems. The formalization of co-evolving belief states, the memory architecture, and the use of uncertainty-aware parsing are interesting and potentially useful contributions. The paper also provides public demos and resources. However, the central empirical claim that SIL achieves 'mutual, bidirectional co-adaptation' is not currently evidenced: the human belief state is never independently measured, and the headline alignment metric is computed from the agent's own internal representations. The significance of the contribution therefore depends on a load-bearing assumption that the paper does not validate.

major comments (4)
  1. [Section III-B, Eq. (4) and Section IV-C] The belief alignment ρ_t is defined as a confidence-weighted cosine similarity between the agent's estimate of the human belief zH_t and the agent's own belief zA_t. However, zH_t is updated in Eq. (3) using the agent's own encoder outputs and the agent-defined influence matrix W_AH. Thus ρ≈0.83 primarily measures the internal consistency of a single model's representations, not the alignment between the human's actual belief and the agent's belief. The paper reports no user study, no self-report data, and no independent behavioral measure of the human's belief state. This is a load-bearing gap because the paper's central claim—'mutual, bidirectional interactions' and 'both the human and the agent maintain joint belief states'—rests on this metric. The authors should either provide independent evidence of human belief (e.g., human judgments or behavioral measures) or substantially refram
  2. [Section IV-A and Table II] The evaluation is conducted with experimenter-designed commands ('we designed task instructions', 'we issued ambiguous instructions'), with no participant data, no trial counts, and no statistical tests. The claimed ~20-point improvement over ablations is presented without variance estimates or significance testing. The error bars in Fig. 4 are mentioned but no number of trials is given, making it impossible to assess whether the differences are reliable. This is especially important because the ablation 'w/o Co-Adaptation' reduces TCR to 61.7%, nearly the level of the static LLM baseline (60.1%), so the contribution of the co-adaptation mechanism hinge on a comparison whose statistical basis is undocumented.
  3. [Section III-B, Eq. (3)] The human belief update is modeled with the rule zH_{t+1} = η4 zH_t + η5 zH_new + η6 (αH·(2−s_t)·δH_infl). This is an ad-hoc assumption that the human adapts according to the agent's influence vector and the failure signal (2−s_t). No data, human subject, or learning theory is provided to support this model. Since the headline co-adaptation results are generated by simulating the human side with this equation, the entire experimental validation is an internal simulation of the agent's own assumed human model. The authors should clearly state that the human is simulated, and ideally validate the human model against real human behavior, or restrict the claims to a simulation study.
  4. [Section IV and hyperparameter settings] The framework introduces many free parameters—mixing coefficients η1..η6, adaptation rates αA, αH, misalignment threshold τ_mis, EWC importance λ, triplet margin m, uncertainty weights αu, βu, retrieval weights ws, wb, softmax temperature τ, task-shift threshold τ_shift, and latent dimension d. The paper sets these to specific values (Section IV) but provides no sensitivity analysis. Given that the reported superiority of SIL over ablations is relatively narrow (about 20 points absolute) and that the model has many interacting components, it is not clear that the results are robust to reasonable variations. At minimum, a sensitivity analysis for the most critical parameters (e.g., η, α, τ_mis) is needed to support the claim that the co-adaptation mechanism, rather than specific hyperparameters, drives the performance.
minor comments (4)
  1. [Section III-B] The cross-reference to Eq. (8) appears in the text after Eq. (5), but the EWC term is not defined until Section III-C. Please make the reference explicit or move the definition earlier.
  2. [Table I] The table header lists 'Metrics EIF MIIR QOR PDS LPL' with values such as 'TCR (%)↑87.36'. The arrow notation is unclear: please clarify whether ↑ denotes 'higher is better' and include units consistently. Also, it would be helpful to report the number of trials per task domain.
  3. [Section IV-D and Fig. 5] The qualitative examples in Fig. 5 are described only in the caption. The main text refers to 'representative qualitative examples' but does not discuss what they illustrate beyond the caption. Please add a few sentences in the text to connect these examples to the quantitative findings.
  4. [Section II-A] The related work section mentions several frameworks, but the comparison with [16] (TCC) is vague: it says 'we employed a similar evaluation approach as TCC' but does not state which elements are shared. Please specify the evaluation protocol that was adopted.

Circularity Check

1 steps flagged · score 7.0 of 10

Belief alignment ρ≈0.83 compares the agent's own latent state with the agent's internal estimate of the human's latent state; no independent human measurement exists, so the headline co-adaptation result reduces to self-consistency of one model.

  1. self definitional [Section III-B, Eqs. (3)-(4); metric used in Section IV-B and reported in Table II/Fig. 3]
    "the agent not only updates its internal representation based on the observed inputs and feedback, but also explicitly reasons over its estimate of the human’s latent state, and vice versa. ... we update the task embeddings for both participants as: ... zH t+1 = η4 zH t + η5 zH new + η6(αH ·(2−s t)·δ H infl) ... we compute a confidence-weighted belief alignment ρ t based on the similarity between the human and agent task embeddings as: ρt = 1 + cos(zH t , zA t ) 2 · c H t · c A t"

    ρt in Eq. (4) is defined as the cosine similarity between zH_t and zA_t. But zH_t is not an independent observation of the human's belief: Eq. (3) updates it using zH_new (the agent's own encoder output) and the agent-defined influence δH_infl = tanh(WAH zA_t), and the text explicitly calls zH the agent's 'estimate of the human's latent state.' Both zH_t and zA_t are therefore outputs of the same learned encoder φ, trained by the triplet loss in Eq. (5) on success-labelled episodes. Section IV-A reports only experimenter-designed/issued commands ('We designed task instructions...', 'we issued ambiguous instructions...'), with no user study, self-report, or behavioral measure of the actual human's belief. Thus ρ≈0.83 is a self-consistency score of one model's internal states, and the 'mutua

full rationale

The paper's task-completion results and ablation comparisons are independent, externally observable behavioral measures and are not circular. However, the central co-adaptation claim is anchored to the belief alignment metric ρ in Eq. (4), and that metric is computed entirely inside the agent: zH_t is the agent's own estimate of the human belief, updated by the agent's encoder and influence dynamics in Eq. (3), while zA_t is the agent's own belief. No independent measurement of the human's actual belief, preference, or adaptation appears in Section IV; the evaluation is based on commands the authors designed and issued. Hence the ρ≈0.83 result, and the 'mutual, bidirectional co-adaptation' conclusion drawn from Fig. 3, reduce by construction to the internal consistency of the agent's latent model. Because the task-completion and ablation evidence retains independent content, the circularity is partial rather than total: score 7. The self-citations ([2], [16], [28]) are methodological and not load-bearing for the formal derivation, so they do not add a separate circularity finding.

Assumptions & free parameters 10 free parameters · 7 assumptions · 2 invented entities

The framework rests on several hand-set hyperparameters and on the assumption that an agent-side latent variable faithfully represents the human's belief. The headline metric ρ is defined using the model's own representations, which is the main circularity burden. The human update equation is an ad hoc model rather than a measured phenomenon.

free parameters (10)
  • mixing coefficients η1..η6 = η1=0.6, η2=0.3, η3=0.1, η4=0.7, η5=0.2, η6=0.1
    Manual tuning in Eq. (3); control update magnitudes and thus convergence/alignment.
  • adaptation rates αA, αH = 0.1, 0.05
    Manual, Section IV; scale influence of success/failure on updates.
  • misalignment threshold τ_mis = 0.6
    Triggers clarification protocol; chosen manually.
  • EWC importance λ = 1000
    Balances plasticity/stability in Eq. (8), chosen manually.
  • triplet margin m = not stated
    Margin in Eq. (5); value unspecified.
  • uncertainty weights αu, βu = not stated ('empirically determined')
    Weights in Eq. (10) empirically chosen; values not reported.
  • retrieval weights ws, wb and softmax temperature τ = ws=0.6, wb=0.4; τ unspecified
    Balance linguistic vs belief retrieval in Eq. (6); temperature not given.
  • task-shift threshold τ_shift = not stated
    Detects distribution shift for EWC checkpoints in Section III-C.
  • latent dimension d = 256
    Dimensionality of shared space, Section IV.
  • influence matrices W_HA, W_AH = learned, no training objective specified
    Eq. (2) transforms partner belief; no loss/update rule given.
assumptions (7)
  • domain assumption A shared d-dimensional latent space can faithfully represent both human and agent task beliefs and language commands.
    Everything (beliefs, commands, memory) is embedded in Z; no evidence this space is sufficient.
  • ad hoc to paper The human's belief evolves according to the agent's update Eq. (3), with failure causing human adaptation via (2−s_t).
    Human side is modeled by the agent; no human-subject data supports this dynamic.
  • ad hoc to paper Cosine similarity weighted by confidence is a valid measure of human-agent belief alignment.
    Eq. (4) defines ρ; used as headline metric without external validation.
  • domain assumption A success signal s_t is available after each interaction.
    Needed for Eq. (3), triplet mining, and evaluation; how success is labeled is not described.
  • standard math EWC diagonal Fisher information adequately protects prior task representations.
    Standard EWC assumption [9].
  • domain assumption Pretrained GPT-4o, CLIP, SAM provide reliable language, perception, and segmentation.
    System relies on their correctness; no failure analysis.
  • domain assumption Monocular depth estimation plus RGB-D is sufficient for grounding 3D coordinates.
    Section III-E; unknown reliability in real scenes.
invented entities (2)
  • Agent-side estimate of human belief zH_t
    purpose: Enables Eq. (3) human update and ρ alignment metric.
    No external measurement of human belief; both ρ and updates use the same encoder ϕ, so high alignment may reflect self-consistency rather than true human-agent agreement.
  • Joint belief state B^H_t / B^A_t in shared latent task space
    purpose: Formalizes co-adaptation and clarification triggers.
    Conceptual construct; not grounded in a measured psychological or behavioral state.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SIL: Symbiotic Interactive Learning for Language-Conditioned Human-Agent Co-Adaptation." pith.science (2026). https://pith.science/paper/KSWOJCYW

@misc{pith2026251105203,
  author       = {Pith},
  title        = {Pith review of: SIL: Symbiotic Interactive Learning for Language-Conditioned Human-Agent Co-Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KSWOJCYW}},
  note         = {Machine review of arXiv:2511.05203}
}
abstract

Today's autonomous agents, largely driven by foundation models (FMs), can understand natural language instructions and solve long-horizon tasks with human-like reasoning. However, current human-robot interaction frameworks largely follow a one-way master-apprentice technique where the embodied agent passively executes commands without reciprocal learning. This neglects the co-adaptive, multi-turn nature of everyday human-to-human interactions. We introduce symbiotic interactive learning (SIL), a bidirectional co-adaptation framework in a shared latent task space, where both the human and the agent maintain joint belief states that evolve with the interaction history. This enables proactive clarification, adaptive suggestions, and shared plan refinement. SIL leverages FMs for spatial perception and reasoning, together with a triplet-loss-trained neural encoder that grounds the FMs' outputs into task-specific latent representations. To support long-term stability as tasks evolve, SIL utilises episodic and semantic memory architectures, regularised via elastic weight consolidation to mitigate catastrophic forgetting. We evaluate SIL on simulated and real-world embodied tasks, including instruction following, information retrieval, query-oriented reasoning, and interactive dialogue, achieving a $90.4\%$ task completion rate and a belief alignment score of $\rho \approx 0.83$, an absolute improvement of about $20$ percentage points over the best ablations. Demos and resources: https://linusnep.github.io/SIL/.

Figures

Figures reproduced from arXiv: 2511.05203 by the authors.

Figure 1
Figure 1. (a) The traditional master-apprentice model places the entire reasoning burden on the user (e.g., context, memory), requiring precise and unambiguous [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of SIL’s architecture. Human instructions are received through the natural-language interaction interface (A) and passed to the LLM [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Belief alignment (ρ) across multi-turn interactions. Full SIL (blue) exhibits rapid convergence toward a stable equilibrium ρ ≈ 0.83, maintaining high alignment throughout. In contrast, ablations without co￾adaptation, EWC, human preference modelling, memory, or uncertainty handling exhibit unstable trajectories (ρ ≈ 0.52 − 0.65) and fail to achieve strong alignment. retained and applied this preference in later com… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Task success rate across domains and ablated variants. Full SIL [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative examples of SIL in multi-turn interaction tasks. Yellow paths indicate the agent’s navigation trajectories, starting from the origin [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 3 linked inside Pith

  1. [1]

    Do as i can, not as i say: Grounding language in robotic affordances,

    A. Brohan, Y . Chebotar,et al., “Do as i can, not as i say: Grounding language in robotic affordances,” inConference on robot learning, pp. 287–318, PMLR, 2023

    2023

  2. [2]

    Reli: A language-agnostic approach to human-robot interaction,

    L. Nwankwo, B. Ellensohn, O. ¨Ozdenizci, and E. Rueckert, “Reli: A language-agnostic approach to human-robot interaction,”arXiv preprint arXiv:2505.01862, 2025

    arXiv 2025

  3. [3]

    When robots get chatty: Grounding multimodal human-robot conversation and collaboration,

    P. Allgeuer, H. Ali, and S. Wermter, “When robots get chatty: Grounding multimodal human-robot conversation and collaboration,” inInternational Conference on Artificial Neural Networks, pp. 306– 321, Springer, 2024

    2024

  4. [4]

    Harnessing the power of llms in practice: A survey on chatgpt and beyond,

    J. Yang, H. Jin,et al., “Harnessing the power of llms in practice: A survey on chatgpt and beyond,”ACM Transactions on Knowledge Discovery from Data, vol. 18, no. 6, pp. 1–32, 2024

    2024

  5. [5]

    Grounding large language models in real-world environments using imperfect world models,

    L. Zhang, Z. Liu, Y . Zhou, T. Wu, and J. Sun, “Grounding large language models in real-world environments using imperfect world models,”IJACSA) International Journal of Advanced Computer Sci- ence and Applications, 2024

    2024

  6. [6]

    Envodat: A large-scale multisen- sory dataset for robotic spatial awareness and semantic reasoning in heterogeneous environments,

    L. Nwankwo, B. Ellensohn, V . Dave, P. Hofer, J. Forstner, M. Vill- neuve, R. Galler, and E. Rueckert, “Envodat: A large-scale multisen- sory dataset for robotic spatial awareness and semantic reasoning in heterogeneous environments,” in2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 153–160, 2025

    2025

  7. [7]

    Kolb’s learning styles and experiential learning cycle,

    S. McLeod, “Kolb’s learning styles and experiential learning cycle,” Simply psychology, vol. 5, 2017

    2017

  8. [8]

    Learning in and about complex systems,

    J. D. Sterman, “Learning in and about complex systems,”System dynamics review, vol. 10, no. 2-3, pp. 291–330, 1994

    1994

Show all 35 references
  1. [9]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu,et al., “Overcoming catastrophic forgetting in neural networks,”Proceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

    2017

  2. [10]

    Gpt-4 technical report,

    OpenAI, J. Achiam, and et al., “Gpt-4 technical report,” 2024

    2024

  3. [11]

    Gemini: a family of highly capable multimodal models,

    G. Team, R. Anil,et al., “Gemini: a family of highly capable multimodal models,”arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  4. [12]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim,et al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning, pp. 8748–8763, PmLR, 2021

    2021

  5. [13]

    Grounded language-image pre-training,

    L. H. Li, P. Zhang,et al., “Grounded language-image pre-training,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10965–10975, 2022

    2022

  6. [14]

    Interactive language: Talking to robots in real time,

    C. Lynch, A. Wahid, J. Tompson, T. Ding, J. Betker, R. Baruch, T. Armstrong, and P. Florence, “Interactive language: Talking to robots in real time,”IEEE Robotics and Automation Letters, 2023

    2023

  7. [15]

    Progprompt: program generation for situated robot task planning using large language models,

    I. Singh, V . Blukis,et al., “Progprompt: program generation for situated robot task planning using large language models,”Autonomous Robots, vol. 47, no. 8, pp. 999–1012, 2023. c a b Fig. 5. Qualitative examples of SIL in multi-turn interaction tasks. Yellow paths indicate th...

    2023

  8. [16]

    The conversation is the command: Interacting with real-world autonomous robots through natural lan- guage,

    L. Nwankwo and E. Rueckert, “The conversation is the command: Interacting with real-world autonomous robots through natural lan- guage,” inCompanion of the 2024 ACM/IEEE International Confer- ence on Human-Robot Interaction, pp. 808–812, 2024

    2024

  9. [17]

    A survey of robot learning from demonstration,

    B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,”Robot. Auton. Syst., vol. 57, p. 469–483, May 2009

    2009

  10. [18]

    Recent advances in robot learning from demonstration,

    H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard, “Recent advances in robot learning from demonstration,”Annual review of control, robotics, and autonomous systems, vol. 3, no. 1, pp. 297– 330, 2020

    2020

  11. [19]

    Interactive policy learning through confidence-based autonomy,

    S. Chernova and M. Veloso, “Interactive policy learning through confidence-based autonomy,”Journal of Artificial Intelligence Re- search, vol. 34, pp. 1–25, 2009

    2009

  12. [20]

    Deep reinforcement learning from human preferences,

    P. F. Christiano, J. Leike,et al., “Deep reinforcement learning from human preferences,”Advances in neural information processing sys- tems, vol. 30, 2017

    2017

  13. [21]

    Humanizing human-robot interaction: On the importance of mutual understanding,

    A. Sciutti, M. Mara,et al., “Humanizing human-robot interaction: On the importance of mutual understanding,”IEEE Technology and Society Magazine, vol. 37, no. 1, pp. 22–29, 2018

    2018

  14. [22]

    Shared autonomy via hindsight optimization,

    S. Javdani, S. S. Srinivasa, and J. A. Bagnell, “Shared autonomy via hindsight optimization,”Robotics science and systems: online proceedings, vol. 2015, pp. 10–15607, 2015

    2015

  15. [23]

    Formalizing human-robot mutual adaptation: A bounded memory model,

    S. Nikolaidis, A. Kuznetsov, D. Hsu, and S. Srinivasa, “Formalizing human-robot mutual adaptation: A bounded memory model,” in2016 11th ACM/IEEE International Conference on Human-Robot Interac- tion (HRI), pp. 75–82, 2016

    2016

  16. [24]

    An effective personal mobile robot agent through symbiotic human-robot interaction.,

    S. Rosenthal, J. Biswas, and M. M. Veloso, “An effective personal mobile robot agent through symbiotic human-robot interaction.,” in AAMAS, vol. 10, pp. 915–922, 2010

    2010

  17. [25]

    Generative expressive robot behaviors using large language models,

    K. Mahadevan, J. Chien,et al., “Generative expressive robot behaviors using large language models,” inProceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, pp. 482–491, 2024

    2024

  18. [26]

    Dialogue learning with human teaching and feedback in end-to-end trainable task-oriented dialogue systems,

    B. Liu, G. Tur, D. Hakkani-Tur, P. Shah, and L. Heck, “Dialogue learning with human teaching and feedback in end-to-end trainable task-oriented dialogue systems,”arXiv preprint arXiv:1804.06512, 2018

    2018 arXiv

  19. [27]

    Toward embodied intelligence-enabled human–robot symbiotic manufacturing: A large language model-based perspective,

    W. Dong, S. Li, and P. Zheng, “Toward embodied intelligence-enabled human–robot symbiotic manufacturing: A large language model-based perspective,”Journal of Computing and Information Science in Engi- neering, vol. 25, no. 5, p. 050801, 2025

    2025

  20. [28]

    Multimodal human-autonomous agents interaction using pre-trained language and visual foundation models,

    L. Nwankwo and E. Rueckert, “Multimodal human-autonomous agents interaction using pre-trained language and visual foundation models,” arXiv preprint arXiv:2403.12273, 2024

    2024 arXiv

  21. [29]

    Bayesian ensemble learning,

    H. Chipman, E. George, and R. McCulloch, “Bayesian ensemble learning,”Advances in neural information processing systems, vol. 19, 2006

    2006

  22. [30]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo,et al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 4015–4026, 2023

    2023

  23. [31]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer,

    R. Ranftl, K. Lasinger,et al., “Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 3, pp. 1623–1637, 2020

    2020

  24. [32]

    Ros: an open-source robot operating system,

    M. Quigley, K. Conley,et al., “Ros: an open-source robot operating system,” inICRA workshop on open source software, vol. 3, p. 5, Kobe, 2009

    2009

  25. [33]

    The marathon 2: A navigation system,

    S. Macenski, F. Martin, R. White, and J. Gin ´es Clavero, “The marathon 2: A navigation system,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020

    2020

  26. [34]

    Improved techniques for grid mapping with rao-blackwellized particle filters,

    G. Grisetti, C. Stachniss, and W. Burgard, “Improved techniques for grid mapping with rao-blackwellized particle filters,”IEEE transac- tions on Robotics, vol. 23, no. 1, pp. 34–46, 2007

    2007

  27. [35]

    Robust monte carlo localization for mobile robots,

    S. Thrun, D. Fox, W. Burgard, and F. Dellaert, “Robust monte carlo localization for mobile robots,”Artificial intelligence, vol. 128, no. 1- 2, pp. 99–141, 2001

    2001

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.