Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

ProVox: Personalization and Proactive Planning for Situated Human-Robot Collaboration

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A short pre-task meta-prompting session lets a language-model planner personalize to a user and proactively propose actions, cutting measured lunch-bag packing time from 18.98 to 11.63 minutes in a within-subjects study.

desk verdict ProVox is a genuine, useful extension of Vocal Sandbox, but the headline 38.7% speedup is inflated because it excludes the meta-prompting setup time that only ProVox requires. read the letter →

arxiv 2506.12248 v1 pith:666XXX2V submitted 2025-06-13 cs.RO cs.AIcs.CLcs.HCcs.LG

Jennifer Grannen , Siddharth Karamcheti , Blake Wulfe , Dorsa Sadigh This is my paper
classification cs.ROcs.AIcs.CLcs.HCcs.LG
keywords personalizationproactiveplanningsituatedhuman-robotcollaborationlargelanguagemodelsmeta-promptingtaskuserstudyhouseholdmanipulation
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 collaborative robots working in a shared physical space should learn each partner's goal and preferred behaviors before the task begins, then use that personalized context to propose helpful next actions instead of waiting for explicit commands. To test this, the authors build ProVox, which adds a meta-prompting interface and a proactive language-model planner to an existing situated-collaboration system. In a within-subjects user study of lunch-bag packing, the system completed three bags in 11.63 minutes on average versus 18.98 minutes for the passive, user-agnostic baseline, and participants rated it higher on ease of use and helpfulness. The paper interprets these results as evidence that personalization and proactivity are both needed for the efficiency and preference gains.

What carries the argument

The central object is the meta-prompted context consisting of a personalized goal statement and a user-taught API of behavior functions, produced through a query-driven graphical interface before any physical interaction. The proactive planner is a language model that, after each executed plan, is prompted to propose an action to perform next and, using the full interaction history and the personalized API, generates a candidate plan that is gated on user confirmation before execution. This pairing lets a single base planner adapt to users with different vocabularies, trust levels, and even conflicting definitions of the same behavior name.

What would settle it

Run the same within-subjects lunch-bag study with a third condition that receives meta-prompting but no proactive suggestions; if that condition matches ProVox's 11.63-minute completion time, then proactive planning is not load-bearing for the speedup. Alternatively, count meta-prompting time in the task clock; if the mean gap to 18.98 minutes ceases to be substantial, the efficiency claim depends on the excluded-setup accounting.

Watch

Extended reading notes

Core claim

ProVox's central claim is that a short pre-interaction meta-prompting session—where the user states their goal in natural language and teaches the robot new behavior functions through a graphical interface—can condition a large-language-model planner well enough that it anticipates the user's intent and proactively proposes plans the user accepts. In the real-robot study, users packed three lunch bags in 11.63±0.9 minutes with ProVox versus 18.98±2.3 minutes with the passive baseline, accepted robot-proposed plans 31.9% of the time, and spent 19.0% less time giving explicit instructions. The component study further shows that full meta-prompted context (personalized goal plus personalized API) yields more helpful proactive suggestions than either partial context or a fixed context, supporting the paper's assertion that personalization is what makes proactive planning useful.

Load-bearing premise

The efficiency claim assumes the 5.58-minute meta-prompting session should not be counted as collaboration time, even though only the ProVox condition incurs it; the measured speedup drops from 38.7% to roughly 9% if that setup time is included.

Editorial extensions

If this is right

  • If the central claim holds, situated robot collaborators should include an explicit upfront personalization phase rather than a fixed global prompt, because the same taught behavior name (e.g., 'pack') maps to different motion decompositions for different users.
  • Gated proactivity—suggesting actions but waiting for confirmation—can cut explicit instruction time by 19.0% while keeping the user in control, suggesting a design principle for assistive robots.
  • The 5.58-minute meta-prompting time is a fixed per-interaction cost that does not scale with task horizon, so the relative efficiency gain should grow on longer or repeated collaborations.
  • The component study implies that neither personalization alone nor proactivity alone is enough: partial context (goal or API only) produced far fewer helpful proactive suggestions than the full meta-prompted context.

Reading between the lines

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

  • If the fixed-cost account of meta-prompting is right, longer-horizon or multi-session collaborations should amplify ProVox's advantage, while very short one-off tasks could plausibly favor the passive baseline once the setup session is counted.
  • The 31.9% acceptance rate for robot-proposed plans could partly reflect users deferring to suggestions to avoid the effort of instructing; a follow-up measuring rejections or modifications would separate efficiency gains from acquiescence.
  • The meta-prompting interface's explicit API inspection and editing could be adapted to other teaching modalities, such as kinesthetic demonstration, which would address the paper's acknowledged reliance on ungrounded language understanding.
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 / 5 minor

Summary. ProVox extends the Vocal Sandbox framework with two contributions: a meta-prompting protocol that lets users specify a personalized goal and teach API functions before interaction, and a proactive language-model planner that suggests next actions gated on user confirmation. The paper reports a component study (N=26) measuring how many proactive suggestions match a reference video under full, partial, or no meta-prompt, and a full-system within-subjects user study (N=9) comparing ProVox to a non-proactive Vocal Sandbox baseline on packing three lunch bags. The headline results are 38.7% faster task completion, higher ease/helpfulness ratings, and reduced instruction time.

Significance. Personalization and proactivity in situated human-robot collaboration are timely and important, and the paper has clear strengths: a transparent within-subjects design, a real-robot full-system evaluation, safety gating of proactive proposals, public code and videos, and a component ablation with 26 participants. If the claims hold, the framework is a useful step beyond passive, user-agnostic systems. However, the principal efficiency claim is sensitive to the decision to exclude 5.58 minutes of setup that only ProVox requires, and the component study's overlap metric is partly built into the input. With reanalysis and reframing, the contributions are likely useful, but the current headline overstates the case.

major comments (4)
  1. [Sec. VI-B] The paper's central claim of '38.7% faster collaborative task performance' is based on task completion times of 11.63±0.9 vs 18.98±2.3 minutes that, as stated, 'exclude the time spent in the meta-prompting interface.' Only the ProVox condition incurs this setup, which averages 5.58±0.9 minutes. Adding it to the ProVox condition gives approximately 17.2 vs 18.98 minutes, roughly 9% faster. Because the task horizon is only three lunch bags, the fixed-cost argument does not make the excluded time negligible. Please report end-to-end times that include setup alongside the interaction-only times, and qualify the abstract and discussion claims accordingly.
  2. [Sec. VI-A] The Meta-Prompt Efficacy measure counts the overlap between the proactively generated plan and the reference plan of the video that participants were asked to reproduce. The Prompted Goal condition contains the user's natural-language description of that same task, and the Prompted API contains functions taught to match the video, so high overlap is partly an artifact of the input content. The Fixed Context baseline necessarily scores low because it is given neither piece of information. This confirms that the prompt contains task-relevant information, but it does not validate that the full-system speedup is caused by personalization plus proactivity. A stronger test would use a held-out task or would evaluate suggestion quality with a goal and API obtained from a different user.
  3. [Sec. VI-B] The full-system comparison conflates the two interventions. ProVox differs from Vocal Sandbox in two ways: it includes pre-task meta-prompting, and it includes proactive suggestions. The component study partially separates these, but it is offline and, as noted above, partly self-referential. Without a real-robot condition that removes only proactivity or only personalization, the statement that 'both meta-prompting and proactivity are critical' (Abstract) is not directly supported by the full-system data.
  4. [Abstract and Sec. VI-B] The abstract and discussion state that ProVox gives '31.9% less user burden,' but the full-system section reports that 'participants accede to robot-proposed plans 31.9% of the time.' An acceptance rate for robot suggestions is not the same as a measured reduction in user burden (e.g., NASA-TLX, instruction count, or perceived effort). Please relabel this metric or collect a direct burden measure.
minor comments (5)
  1. [Sec. IV-A and IV-B] Section IV-A states that 'we do not track the entire interaction history h_t' for plan generation, while Section IV-B says the proactive planner uses 'the full interaction history.' Please reconcile these two statements.
  2. [Sec. I and VI-A] The introduction says the meta-prompting protocol is evaluated 'relative to existing meta-prompting approaches,' but the component study compares only against internal ablations (Fixed Context, Fixed API, Fixed Goal). Please either add a comparison to an existing method or revise the claim.
  3. [Sec. VI-B] The 19.0% reduction in explicit instruction time is presented without a significance test. With N=9, please report a test result or explicitly mark this as a descriptive finding.
  4. [Sec. VI] Statistical reporting would be improved by exact p-values, effect sizes, and confidence intervals rather than only 'p≤0.05,' and by addressing multiple comparisons across the seven Likert-scale items.
  5. [Sec. VI-B] The fixed-cost argument would be more convincing with a break-even analysis: at what task horizon does the 5.58-minute setup become negligible relative to the time savings? The current three-bag task is too short to make this point by inspection.

Circularity Check

1 steps flagged · score 4.0 of 10

Component-study efficacy metric is partly circular because the user-built goal and API already encode the reference plan; the full-system comparison remains an external result.

  1. fitted input called prediction [Section VI-A, "Dependent Measures" and "Participants and Procedure"]
    "We compare ppro with the ground truth task plan pref depicted in the initial demonstration, reporting the count of overlapping function invocations (individual behaviors represented in the reference video) as Meta-Prompt Efficacy. Large overlaps indicate that the proactive planner and the corresponding meta-prompting method are effective in suggesting helpful actions that mimic the ground truth task plan pref."

    In this component study, the conditioning input is not independent of the target. Participants were shown a 26-second video of a robot bagging three specific items and were "tasked with reproducing the behavior in as few language instructions as possible," using the meta-prompting interface to build a personalized goal and API. That user-built goal names the same objects and the taught API encodes the same primitive sequences that appear in the reference plan pref. The paper's efficacy metric then scores the LM planner by how much of the proposed plan overlaps with pref. Thus the target of the evaluation is largely contained in the input context by construction: the planner is asked to re-emit content that the participant explicitly placed in the prompt and API.

full rationale

The paper's central full-system claim is not circular: ProVox is compared head-to-head against a non-proactive Vocal Sandbox baseline in a real physical lunch-bag packing task, with task completion times and user ratings measured from actual interaction. That comparison is external evidence and does not reduce to the method's inputs. The 38.7% speedup excludes the 5.58-minute meta-prompting setup time, which is a legitimate cost-accounting concern about whether the headline is end-to-end, but it is not a circularity: re-including the setup would change the magnitude of the result without making the comparison equivalent to its assumptions. The heavy reliance on the authors' prior Vocal Sandbox work is normal build-on-prior-work and is not load-bearing in a circular sense; the baseline is a distinct system and the new proactivity/personalization components are evaluated against it. The main circularity concern is isolated to the component study in Section VI-A, where the reference plan is seeded into the user-generated goal and API before the LM is asked to reproduce that same plan. This weakens the component-level evidence for "helpful proactive suggestions," but it does not make the full-system outcome tautological. Overall, the derivation chain is mostly self-contained, with one partially self-referential evaluation metric, so a moderate score of 4 is appropriate.

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

No numeric free parameters are fitted to data; the framework is an empirical system built on an external LLM. The most important unstated load is the decision to exclude meta-prompting setup time from the primary speedup, plus reliance on LLM commonsense reasoning, user-taught API fidelity, and self-reported ratings.

assumptions (4)
  • domain assumption LLM commonsense priors are sufficient for generating helpful proactive plans from a personalized prompt.
    Section IV-B relies on GPT-4 Turbo to propose sensible next actions; the paper acknowledges failure is possible and falls back to user commands. The user study is the test of this assumption.
  • domain assumption Behaviors users define through the meta-prompting interface accurately capture their intended robot actions.
    Section IV-A assumes the goal text and taught API functions generalize to the physical interaction. The N=26 study only measures whether the proactive planner's suggestions overlap a reference plan, not whether the taught functions are correct in execution.
  • ad hoc to paper Excluding meta-prompting setup time from task completion is a fair comparison.
    Section VI-B excludes 5.58 minutes of setup time that only ProVox requires; this assumption materially affects the headline 38.7% improvement, dropping it to about 9% if included.
  • domain assumption Self-reported Likert ratings reflect collaboration quality.
    The qualitative results in Section VI-B rely on participant self-report; several subjective differences are not statistically significant with N=9.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProVox: Personalization and Proactive Planning for Situated Human-Robot Collaboration." pith.science (2026). https://pith.science/paper/666XXX2V

@misc{pith2026250612248,
  author       = {Pith},
  title        = {Pith review of: ProVox: Personalization and Proactive Planning for Situated Human-Robot Collaboration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/666XXX2V}},
  note         = {Machine review of arXiv:2506.12248}
}
read the original abstract

Collaborative robots must quickly adapt to their partner's intent and preferences to proactively identify helpful actions. This is especially true in situated settings where human partners can continually teach robots new high-level behaviors, visual concepts, and physical skills (e.g., through demonstration), growing the robot's capabilities as the human-robot pair work together to accomplish diverse tasks. In this work, we argue that robots should be able to infer their partner's goals from early interactions and use this information to proactively plan behaviors ahead of explicit instructions from the user. Building from the strong commonsense priors and steerability of large language models, we introduce ProVox ("Proactive Voice"), a novel framework that enables robots to efficiently personalize and adapt to individual collaborators. We design a meta-prompting protocol that empowers users to communicate their distinct preferences, intent, and expected robot behaviors ahead of starting a physical interaction. ProVox then uses the personalized prompt to condition a proactive language model task planner that anticipates a user's intent from the current interaction context and robot capabilities to suggest helpful actions; in doing so, we alleviate user burden, minimizing the amount of time partners spend explicitly instructing and supervising the robot. We evaluate ProVox through user studies grounded in household manipulation tasks (e.g., assembling lunch bags) that measure the efficiency of the collaboration, as well as features such as perceived helpfulness, ease of use, and reliability. Our analysis suggests that both meta-prompting and proactivity are critical, resulting in 38.7% faster task completion times and 31.9% less user burden relative to non-active baselines. Supplementary material, code, and videos can be found at https://provox-2025.github.io.

Figures

Figures reproduced from arXiv: 2506.12248 by the authors.

Figure 1
Figure 1. We present ProVox (“Proactive Voice”), a framework for personaliza￾tion and proactive planning in the context of situated human-robot collabora￾tions. In the first phase of a collaboration [Top], a human communicates their goals and distinct preferences, enabling the robot to personalize. Throughout the rest of the collaboration [Bottom], the robot continues to incorporate and anticipate their partner’s intent to pr… view at source ↗
Figure 2
Figure 2. ProVox Motivating Example. Existing frameworks for situated human-robot collaboration tend to assume static, hard-coded APIs to inform task planning (gray) that cannot be adapted to new individuals with distinct objectives and preferences. Instead, ProVox allows users to provide high-level goals [Top] and define task-relevant actions (e.g., pack on the [Left]). This enables personalization to a user’s specific vocab… view at source ↗
Figure 3
Figure 3. Meta-Prompting Protocol & Proactive Planning. ProVox develops a novel meta-prompting protocol to collect two critical pieces of information from an individual: their specific goal, and an API of useful behaviors. Crucially, each user has a distinct set of preferences, yielding different goals and behaviors. The female user [Top-Left] wants her children’s lunch to contain Skittles and Rice-Krispies, and teaches the r… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Meta-Prompting User Study Results. We present results evaluating the efficacy of ProVox’s meta-prompt protocol through a N = 26 user study, as well as visualize the diversity of user-defined behaviors. We report the mean and standard error values of the number of helpf…
Figure 5
Figure 5. Figure 5: Full System User Study Results. We present quantitative and qualitative results from our real-world user study of collaborative lunch bag packing. We report the average and standard error task completion times for users collaborating with a ProVox or a Vocal Sandbox sy…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CoRI: Communication of Robot Intent for Physical Human-Robot Interaction

    cs.RO 2025-05 conditional novelty 6.0 of 10

    CoRI automatically generates natural-language intent statements from a robot's image view and planned trajectory, and its user study shows it outperforms scripted and no-communication baselines in clarity.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Collaboration in human-robot teams,

    G. Hoffman and C. L. Breazeal, “Collaboration in human-robot teams,” AIAA 1st Intelligent Systems Technical Conf., 2004

    work page 2004

  2. [2]

    Progress and prospects of the human–robot collaboration,

    A. Ajoudani, A. M. Zanchettin, S. Ivaldi, A. O. Albu-Schäffer, K. Ko- suge, and O. Khatib, “Progress and prospects of the human–robot collaboration,”Autonomous Robots, vol. 42, pp. 957 – 975, 2017

    work page 2017

  3. [3]

    Situated human–robot collaboration: predicting intent from grounded natural language,

    J. Brawer, O. Mangin, A. Roncone, S. Widder, and B. Scassellati, “Situated human–robot collaboration: predicting intent from grounded natural language,”Int. Conf. on Intelligent Robots and Systems, 2018

    work page 2018

  4. [4]

    In-mouth robotic bite transfer with visual and haptic sensing,

    L. Shaikewitzet al., “In-mouth robotic bite transfer with visual and haptic sensing,” inProc. IEEE Int. Conf. Robotics and Automation, 2023

    work page 2023

  5. [5]

    Cooperation of human and machines in assembly lines,

    J. Krüger, T. Lien, and A. Verl, “Cooperation of human and machines in assembly lines,”CIRP Annals, vol. 58, no. 2, pp. 628–646, 2009

    work page 2009

  6. [6]

    A survey on applications of human-robot interaction,

    A.-N. Sharkawy, “A survey on applications of human-robot interaction,” Sensors & Transducers, vol. 251, no. 4, pp. 19–27, 2021

    work page 2021

  7. [7]

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

    M. Ahnet al., “Do as i can and not as i say: Grounding language in robotic affordances,” inConf. on Robot Learning, 2022

    work page 2022

  8. [8]

    Mosaic: A modular system for assistive and interactive cooking,

    H. Wanget al., “Mosaic: A modular system for assistive and interactive cooking,” inConf. on Robot Learning, 2024

    work page 2024

Show all 41 references
  1. [9]

    V ocal Sandbox: Continual learning and adaptation for situated human- robot collaboration,

    J. Grannen, S. Karamcheti, S. Mirchandani, P. Liang, and D. Sadigh, “V ocal Sandbox: Continual learning and adaptation for situated human- robot collaboration,” inConf. on Robot Learning, 2024

    2024

  2. [10]

    GPT-4 technical report,

    OpenAI, “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  3. [11]

    Code as policies: Language model programs for embodied control,

    J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng, “Code as policies: Language model programs for embodied control,” inProc. IEEE Int. Conf. Robotics and Automation, 2023

    2023

  4. [12]

    Inner monologue: Embodied reasoning through planning with language models,

    W. Huanget al., “Inner monologue: Embodied reasoning through planning with language models,” inConf. on Robot Learning, 2022

    2022

  5. [13]

    Robots that use language,

    S. Tellexet al., “Robots that use language,”Annual Review of Control, Robotics, and Autonomous Systems, 2020

    2020

  6. [14]

    RT-2: Vision-language-action models transfer web knowledge to robotic control,

    A. Brohanet al., “RT-2: Vision-language-action models transfer web knowledge to robotic control,”arXiv preprint arXiv:2307.15818, 2023

    2023 arXiv

  7. [15]

    Learning from unscripted deictic gesture and language for human-robot interactions

    C. Matuszek, L. Bo, L. Zettlemoyer, and D. Fox, “Learning from unscripted deictic gesture and language for human-robot interactions.” in Proceedings of the AAAI Conf. on Artificial Intelligence, vol. 28, 2014

    2014

  8. [16]

    Gesture-informed robot assistance via foundation models,

    L.-H. Linet al., “Gesture-informed robot assistance via foundation models,” inConf. on Robot Learning, 2023

    2023

  9. [17]

    "No, to the right

    Y . Cuiet al., “"No, to the right"– online language corrections for robotic manipulation via shared autonomy,” inACM/IEEE HRI, 2023

    2023

  10. [18]

    Yell at your robot: Improving on-the-fly from language corrections,

    L. X. Shiet al., “Yell at your robot: Improving on-the-fly from language corrections,”ArXiv, vol. abs/2403.12910, 2024

    2024 arXiv

  11. [19]

    Learning robot objectives from physical human interaction,

    A. Bajcsyet al., “Learning robot objectives from physical human interaction,” inConf. on Robot Learning, 2017

    2017

  12. [20]

    Progprompt: Generating situated robot task plans using large language models,

    I. Singhet al., “Progprompt: Generating situated robot task plans using large language models,”IEEE Int. Conf. Robotics and Automation, 2022

    2022

  13. [21]

    Ok-robot: What really matters in integrating open-knowledge models for robotics,

    P. Liu, Y . Orru, C. Paxton, N. M. M. Shafiullah, and L. Pinto, “Ok-robot: What really matters in integrating open-knowledge models for robotics,” inProc. Robotics: Science and Systems, 2024

    2024

  14. [22]

    Embodied ai with two arms: Zero-shot learning, safety and modularity,

    J. Varleyet al., “Embodied ai with two arms: Zero-shot learning, safety and modularity,”arXiv preprint arXiv:2404.03570, 2024

    2024 arXiv

  15. [23]

    Grounded decoding: Guiding text generation with grounded models for robot control,

    W. Huanget al., “Grounded decoding: Guiding text generation with grounded models for robot control,” inProc. Advances in Neural Information Processing Systems, 2023

    2023

  16. [24]

    How to prompt your robot: A promptbook for manipulation skills with code as policies,

    M. G. Arenaset al., “How to prompt your robot: A promptbook for manipulation skills with code as policies,”IEEE Int. Conf. on Robotics and Automation, pp. 4340–4348, 2024

    2024

  17. [25]

    GPT-3.5 – Function calling and other updates,

    OpenAI, “GPT-3.5 – Function calling and other updates,” https://openai. com/index/function-calling-and-other-api-updates/, 2023

    2023

  18. [26]

    Naturalizing a programming language via interactive learning,

    S. I. Wanget al., “Naturalizing a programming language via interactive learning,” inAssociation for Computational Linguistics, 2017

    2017

  19. [27]

    Learning preferences for manipulation tasks from online coactive feedback,

    A. Jainet al., “Learning preferences for manipulation tasks from online coactive feedback,”Int. Journal of Robotics Research, 2015

    2015

  20. [28]

    Batch active preference-based learning of reward functions,

    E. Biyik and D. Sadigh, “Batch active preference-based learning of reward functions,” inConf. on Robot Learning, 2018

    2018

  21. [29]

    Aprel: A library for active preference-based reward learning algorithms,

    E. Biyik, A. Talati, and D. Sadigh, “Aprel: A library for active preference-based reward learning algorithms,”ACM/IEEE HRI, 2021

    2021

  22. [30]

    Efficient model learning from joint-action demon- strations for human-robot collaborative tasks,

    S. Nikolaidiset al., “Efficient model learning from joint-action demon- strations for human-robot collaborative tasks,”ACM/IEEE HRI, 2014

    2014

  23. [31]

    Information gathering actions over human internal state,

    D. Sadighet al., “Information gathering actions over human internal state,”IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, 2016

    2016

  24. [32]

    Shared autonomy via hindsight optimization for teleoperation and teaming,

    S. Javdaniet al., “Shared autonomy via hindsight optimization for teleoperation and teaming,”Int. Journal of Robotics Research, 2018

    2018

  25. [33]

    Learning adaptive language interfaces through decomposition,

    S. Karamcheti, D. Sadigh, and P. Liang, “Learning adaptive language interfaces through decomposition,” inEMNLP Workshop for Interactive and Executable Semantic Parsing, 2020

    2020

  26. [34]

    Test driven development: By example,

    Beck, “Test driven development: By example,” 2002. [Online]. Available: https://api.semanticscholar.org/CorpusID:262220275

    2002

  27. [35]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing,

    P. Liuet al., “Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing,”ACM Computing Surveys, vol. 55, pp. 1 – 35, 2021

    2021

  28. [36]

    Human-robot teaming using shared mental models,

    S. Nikolaidis and J. Shah, “Human-robot teaming using shared mental models,”ACM/IEEE HRI, 2012

    2012

  29. [37]

    DROID: A large-scale in-the-wild robot manipu- lation dataset,

    A. Khazatskyet al., “DROID: A large-scale in-the-wild robot manipu- lation dataset,” inRobotics: Science and Systems, 2024

    2024

  30. [38]

    Fast segment anything,

    X. Zhao, W. Ding, Y . An, Y . Du, T. Yu, M. Li, M. Tang, and J. Wang, “Fast segment anything,”arXiv preprint arXiv:2306.12156, 2023

    2023 arXiv

  31. [39]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved baselines with visual instruction tuning,”arXiv preprint arXiv:2310.03744, 2023

    2023 arXiv

  32. [40]

    GPT-4v(ision) system card,

    OpenAI, “GPT-4v(ision) system card,” 2023

    2023

  33. [41]

    Openvla: An open-source vision-language-action model,

    M. J. Kimet al., “Openvla: An open-source vision-language-action model,” inConf. on Robot Learning, 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.