Pith. sign in

REVIEW 4 major objections 6 minor 32 references

PSALM-V: Automating Symbolic Planning in Interactive Visual Environments with Large Language Models

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PSALM-V claims that an autonomous vision-language loop can learn symbolic action rules in interactive environments, lifting plan success from 37% to 74%.

desk verdict PSALM-V has the right core idea for learning PDDL semantics in visual environments, but the printed belief-update rule is self-defeating as written and the empirical claims need tighter reporting before the results can be taken at face value. read the letter →

arxiv 2506.20097 v1 pith:K4DPZNMS submitted 2025-06-25 cs.RO cs.CL

Wang Bill Zhu , Miaosen Chai , Ishika Singh , Robin Jia , Jesse Thomason This is my paper
classification cs.ROcs.CL
keywords symbolicplanningPDDLdomaininductionneuro-symboliclearningvision-languagemodelspartialobservabilityactionsemanticsinteractiveenvironmentsroboticmanipulation
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

PSALM-V argues that symbolic planners no longer need hand-written action rules: the system predicts PDDL problem files and action semantics from visual observations, executes candidate plans, explains failures, and revises its beliefs until a goal is reached. The point of doing this in a loop is that an LLM alone can propose plausible plans but is bad at tracking state and correcting itself, while a symbolic planner is reliable only if given the rules. PSALM-V claims to supply those rules by induction from interaction, and to do so in partially observed, visually grounded environments with no expert action definitions, no problem file, and no explicit error messages. The paper reports that this increases ALFRED plan success from 37% to 74% over a strong LLM baseline and recovers complete BlocksWorld domains on a real robot in about 9.3 environment steps.

What carries the argument

The load-bearing mechanism is a tree-structured belief over the pre- and post-condition formulas of each action. Each tree has logical connector nodes ($and$, $or$, $when$) and statement nodes $\phi$ with belief $p(\phi|a)$; the negation $\neg\phi$ is a sibling whose belief is reduced whenever $\phi$ is predicted. After each execution, beliefs are updated with a contradiction penalty $\alpha=0.7$ and an exponential forgetting factor $\beta=0.8$, then a domain file is sampled from memory and handed to a symbolic planner with a 30-second search budget. If no plan is found, the partial trajectory and predicted error explanation feed the next sample-execute-explain iteration. A separate prospection step symbolically checks the first five actions of each proposed trajectory against current preconditions before execution.

What would settle it

Run PSALM-V in a visual environment whose goal requires a predicate that was deliberately omitted from the supplied vocabulary; the paper's own limitation predicts the induction loop will fail or stall. If the system still recovers the domain and reaches the goal, the stated vocabulary dependence is false; if it fails, the scope of the central claim is confirmed.

Watch

Extended reading notes

Core claim

The central claim is that preconditions and postconditions of actions can be induced autonomously in visual environments by treating them as the target of an iterative belief-refinement loop. Starting from a problem file generated from visual input, PSALM-V samples trajectories with an LLM, executes them, predicts what went wrong (unsatisfied preconditions or unreached goals), and uses that feedback to update a tree-structured belief over each action's semantics; a symbolic planner then validates whether the current belief supports a path to the goal. The paper reports full domain recovery (F1 100%) in RTFM and Overcooked-AI, a 74% success rate and 75% goal-conditioned completion on unseen ALFRED tasks under partial observability, and complete BlocksWorld action semantics on a physical robot with 9.3 average environment steps, even when low-level manipulation errors occur.

Load-bearing premise

PSALM-V succeeds only if the supplied vocabulary of action names, object types, and predicates is correct and complete, because it learns which rules connect those symbols but never discovers what the basic symbols themselves should be.

Editorial extensions

If this is right

  • In ALFRED's unseen validation split, PSALM-V lifts plan success from 37% with a strong off-the-shelf LLM to 74%, and goal-conditioned completion from 57% to 75%.
  • In RTFM and Overcooked-AI, PSALM-V attains 100% win rates and F1 100% domain induction, including in multi-agent settings.
  • On a real robot, PSALM-V recovers the full BlocksWorld domain with F1 100% in about 9.3 average environment steps, reaching 66.7% goal-conditioned completion even when low-level manipulation fails.
  • Every component is load-bearing: removing error-message prediction drops ALFRED domain F1 from 91 to 64, and disabling problem-file checking roughly doubles or worse the executed steps needed.
  • The induced action semantics support $and$, $or$, and $when$ connectors, going beyond prior work that restricted semantics to conjunctions.

Reading between the lines

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

  • The authors' stated limitation points to the natural next step: let the system propose candidate predicates and action signatures from visual affordances; if that succeeded, the remaining human input would disappear entirely.
  • Because the ablations show that unreliable error explanation and problem-file checking inflate execution cost far more than planning search does, making LLM-generated feedback self-verifying through constrained decoding or multi-run consistency checks may be the highest-leverage improvement for public-model versions.
  • The learned domain files are reusable world models: once PSALM-V induces a domain, that symbolic description could be cached and reapplied to new problems in the same environment, turning one expensive induction episode into cheap planning later.
  • The real-robot result suggests the method can tolerate low-level actuator noise because the learner distinguishes symbolic action failure from manipulation failure; an explicit test would be to increase manipulation error rates and measure whether domain induction F1 remains at 100% or begins to degrade.
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 / 6 minor

Summary. PSALM-V proposes an iterative neuro-symbolic framework for inducing PDDL action pre- and post-conditions in partially observed, visually grounded environments. The system initializes a PDDL problem file from visual observations, samples LLM-generated trajectories, executes them in the environment, predicts error messages on failure, and maintains a tree-structured belief over candidate action semantics that is used to sample domain files for a symbolic planner (Fast Downward). This loop repeats until a plan that reaches the goal is found. The paper reports plan success improvements on ALFRED (from 37% for Claude-3.7 direct planning to 74% for PSALM-V), 100% win rates on RTFM and Overcooked-AI, and full domain recovery (F1=100%) on three real-robot BlocksWorld tasks despite low-level manipulation failures. The central claim is that PSALM-V is the first autonomous neuro-symbolic system to induce symbolic action semantics in visual environments through interaction.

Significance. If the mechanism works as presented, the contribution is potentially significant: it removes the need for hand-written pre/post-condition rules, uses execution feedback rather than explicit error messages, and couples LLM heuristics with symbolic verification, which is a natural and valuable direction. The ablations degrade in sensible directions (prospection, error messages, problem-file checking, and RAG all matter), and the real-robot demonstration, though small, shows robustness to low-level execution noise. The paper also includes a candid limitation section. However, the significance is conditional: the central belief-update rule is printed with a sign error that makes the learning loop impossible as written, and the evaluation has several rigor gaps, including a modified RTFM setting and a redefined win-rate metric. With those issues resolved, this could be a solid contribution to LLM-based planning and domain induction.

major comments (4)
  1. [Section 3, Tree-structured memory update] The displayed update rule uses min(..., 0) in both branches, so p_{t+1}(φ|a) is always ≤ 0. For a newly predicted statement with no negation, the first branch evaluates to min(1, 0) = 0, so a node that is first predicted never acquires positive belief. In the second branch, a positive existing belief is also clamped to 0 whenever the interpolated value is positive. This directly contradicts the prose two sentences earlier, which states that a node first predicted is initialized to 1, and it makes the reported learning loop impossible: no new pre- or post-condition can ever become believed. If the implementation instead uses max(·, 0) or a different clamping scheme, the printed formula is a sign error; if it literally uses min, the method cannot function as claimed. Since no code is released, this is an internal inconsistency at the central algorithmic step. The authors must correct the formula and provide a precise, executable pseudocode or code to confirm the intended update.
  2. [Section 4 and Appendix C.1] The RTFM evaluation is not comparable to the original benchmark: the paper states that enemy movement is disabled, and Appendix C.1 redefines the win rate as 'the winning score over optimal winning score in 100 steps.' Because RTFM is used as evidence for domain induction and multi-agent planning, the authors should report the standard unmodified win rate as well, and clarify how the static-enemy setting affects the difficulty. In addition, no number of evaluation episodes or random seeds is given for RTFM or Overcooked-AI, so the reported 100% figures cannot be distinguished from point estimates on a small sample.
  3. [Section 6 Limitation and Section 3] The manuscript's central framing says PSALM-V works 'without expert action definitions,' but §3 states that the input includes 'a list of domain action signatures—a string name for the action and the type and number of arguments it takes'—and §6 explicitly acknowledges that predefined action names, object types, and predicates are required. Thus the system induces only the pre/post-condition structure over a supplied vocabulary; it does not discover action signatures or predicates. This is disclosed in the limitation section, but the abstract and introduction should be reworded to avoid implying full autonomy. The 'first autonomous neuro-symbolic learning system' claim should be scoped to 'first to induce action semantics given a fixed symbolic vocabulary in visual environments.'
  4. [Section 5 and Appendix A.2] The real-robot evaluation consists of only three hand-selected BlocksWorld tasks, with no reported trials per task or error bars, and the problem-file accuracy is described as accurate 'from manual inspection' (Appendix A.2). The 100% F1 and 66.7% goal-conditioned completion are encouraging, but they are not sufficient to support the strength of the claims. The authors should report per-task repetitions, stochastic variation, and a baseline that has access to the same action signatures and predicates but does not perform semantic learning, to isolate the contribution of the induction loop.
minor comments (6)
  1. [Table 4 and Figure 4] The ablation results are reported as single point estimates without confidence intervals or significance tests. Given the stochasticity of LLM-based components, the claim that 'all components contribute to accuracy and step efficiency' would be much stronger with multiple runs and variance reporting.
  2. [Section 3 and Appendix E] The prompt for the error message predictor (Figure 15) appears to be identical to the prompt for the problem file checker (Figure 14), even though the text describes two different roles. Please ensure the prompts are actually distinct and describe how the error message predictor's output differs from the problem file checker's output.
  3. [Abstract and Table 2] The comparison 'from 37% (Claude-3.7) to 74%' mixes two different protocols: Claude-3.7 is direct closed-loop LLM planning, while PSALM-V is an iterative system that can reset and learn. This should be stated explicitly when the number is quoted, since the success rate alone overstates the improvement if the reader assumes a single-attempt comparison.
  4. [Appendix C.2] The prospection mechanism is described as rejecting actions that violate preconditions until a valid prefix is found, but there is no stated halting condition or maximum retry count. Please specify what happens if the LLM repeatedly fails to produce a valid prefix.
  5. [Table 3] The row for 'GPT-4o -Plan' has misaligned columns (the F1, GC, and NES values appear to be shifted across tasks), making the table hard to read. Please reformat the table and clearly indicate which entries are undefined.
  6. [Appendix C.1] There is a typo in 'Overcooked-AI is evluated on the win rate.' Also, the exact number of evaluation episodes for each environment should be reported in the main text or appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned semantics are validated against external execution and ground-truth domains, and the self-citations are not load-bearing.

full rationale

PSALM-V's chain is: vision -> problem-file initialization -> LLM trajectory sampling -> environment execution -> error-message prediction -> action-semantics generation -> tree memory update -> symbolic planner verification. None of these steps defines the target from itself: pre/post-conditions are scored against ground-truth domain files (F1) or environment-defined success (ALFRED SR/GC, RTFM/Overcooked win rate), and the real-robot domain is recovered against the known BlocksWorld semantics. The required vocabulary (action names, object types, predicates) is an input scope limitation, acknowledged in the paper's limitation section, but the paper's claim is specifically the induction of pre/post-conditions, not predicate discovery, so the limitation does not make the result definitionally circular. The citations to the authors' own prior work (Zhu et al. [2]) only motivate that error messages are useful for semantic inference; that premise is independently supported by the paper's own ablation (F1 drops from 91 to 64 when error messages are removed) and by the external execution results, so it is not load-bearing. The same-LLM plan/error loop is an intermediate self-diagnosis step, not a fitted target: the eventual evaluation is execution-based and ground-truth-based, so it does not reduce to its inputs by construction. Separately, the printed update rule in Section 3 uses min(...,0), which would force newly suggested conditions to zero belief and prevent the claimed refinement; this is an internal-consistency/correctness defect (likely a sign error), not a circular reduction, and it does not affect the circularity score. Appendix C.1's statement that ALFRED has no ground truth domain file while Table 2 reports F1 is likewise a reporting gap, not circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical entities are postulated; the tree-structured belief memory is a data structure, not an invented entity. The central claim rests on five hand-chosen scalars (alpha, beta, k, W, RAG count) and on six assumptions, the most load-bearing being the supplied symbolic vocabulary, which the authors themselves flag. The static-enemy RTFM modification is an ad hoc evaluation assumption that weakens the generality of the 100% win-rate claim.

free parameters (5)
  • contradiction penalty alpha = 0.7
    Hand-chosen in the belief update rule (§3, Tree-structured memory update). No sensitivity analysis is reported.
  • forgetting factor beta = 0.8
    Hand-chosen decay factor for prior belief; no sensitivity analysis is reported (§3).
  • prospection window k = 5
    Hand-chosen number of steps checked symbolically before execution (§3 and Appendix C.2).
  • planner search time limit W = 30 seconds
    FAST-DOWNWARD time limit per PSALM-V loop (§3).
  • RAG retrieval count = top-2
    Number of training-set instructions retrieved for ALFRED goal generation (§3).
assumptions (6)
  • standard math The symbolic PDDL model is deterministic and fully observable within the planner's state representation
    Classical planning assumptions inherited from PDDL and FAST-DOWNWARD (§2.1); low-level execution noise is handled outside the symbolic model.
  • domain assumption Predefined action names, object types, and predicates are correct and sufficient
    Admitted in §6 Limitation: PSALM-V requires these to induce action semantics and cannot propose new predicates or action signatures itself.
  • domain assumption VLM and semantic-map translation from images to PDDL initial states is approximately correct
    Section 3, Problem file initialization; the separate problem-file checker exists precisely because this translation is imperfect, so recovering from errors is part of the design.
  • domain assumption LLM-generated error messages are informative enough to drive semantic updates
    Section 3; the ablation without error messages drops ALFRED F1 from 91 to 64, and Qwen-2.5-72B replacements degrade further (Appendix D.3).
  • ad hoc to paper Static-enemy RTFM remains a valid test of partial observability and planning
    Appendix C.1 states 'we disable the movement of enemies' to make symbolic planning applicable, and the win rate is redefined as winning score over optimal in 100 steps.
  • domain assumption BlocksWorld action semantics are representative of the domain-induction problem
    Appendix A.1 notes GPT-4o can already generate correct plans in these 4-5 block tasks, so the real-robot test mostly measures verification and recovery from low-level failures rather than difficult induction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PSALM-V: Automating Symbolic Planning in Interactive Visual Environments with Large Language Models." pith.science (2026). https://pith.science/paper/K4DPZNMS

@misc{pith2026250620097,
  author       = {Pith},
  title        = {Pith review of: PSALM-V: Automating Symbolic Planning in Interactive Visual Environments with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4DPZNMS}},
  note         = {Machine review of arXiv:2506.20097}
}
read the original abstract

We propose PSALM-V, the first autonomous neuro-symbolic learning system able to induce symbolic action semantics (i.e., pre- and post-conditions) in visual environments through interaction. PSALM-V bootstraps reliable symbolic planning without expert action definitions, using LLMs to generate heuristic plans and candidate symbolic semantics. Previous work has explored using large language models to generate action semantics for Planning Domain Definition Language (PDDL)-based symbolic planners. However, these approaches have primarily focused on text-based domains or relied on unrealistic assumptions, such as access to a predefined problem file, full observability, or explicit error messages. By contrast, PSALM-V dynamically infers PDDL problem files and domain action semantics by analyzing execution outcomes and synthesizing possible error explanations. The system iteratively generates and executes plans while maintaining a tree-structured belief over possible action semantics for each action, iteratively refining these beliefs until a goal state is reached. Simulated experiments of task completion in ALFRED demonstrate that PSALM-V increases the plan success rate from 37% (Claude-3.7) to 74% in partially observed setups. Results on two 2D game environments, RTFM and Overcooked-AI, show that PSALM-V improves step efficiency and succeeds in domain induction in multi-agent settings. PSALM-V correctly induces PDDL pre- and post-conditions for real-world robot BlocksWorld tasks, despite low-level manipulation failures from the robot.

Figures

Figures reproduced from arXiv: 2506.20097 by the authors.

Figure 1
Figure 1. LLMs can propose plans and generate action semantics, but struggle with state tracking. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Starting with a problem file initialized from visual input, PSALM-V iteratively samples and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Problem file initialization in the partially observed environment. We first explore the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Removing PSALM-V components such as prospection, error message prediction, and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: PSALM-V remains robust even in the presence of low-level execution failures. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The initial state of robot BlockWorld task 1 on 4 blocks [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: The initial state of robot BlockWorld task 2 on 5 blocks [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: The initial state of robot BlockWorld task 3 on 5 blocks [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Simulated envs: (left) ALFRED [4], (middle) RTFM [5], (right) Overcooked-AI [6]. (stack green blue)), and the state transition from the initial state of the action and the end state of the action (i.e., two images) to verify if the state transition matches the action d…
Figure 10
Figure 10. Figure 10: Prompt for PDDL problem file generation. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Prompt for PDDL goal generation. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Prompt for trajectory sampling. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Prompt for action semantics generator. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Prompt for problem file checker. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]
Figure 15
Figure 15. Figure 15: Prompt for error message predictor. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 28 canonical work pages

  1. [1]

    Ghallab, A

    M. Ghallab, A. Howe, C. Knoblock, D. McDermott, A. Ram, M. Veloso, D. Weld, and D. Wilkins. PDDL - the planning domain definition language.Technical Report, Tech. Rep., 1998

    work page 1998

  2. [2]

    W. Zhu, I. Singh, R. Jia, and J. Thomason. Language models can infer action semantics for symbolic planners from environment feedback. InProceedings of the 2025 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2024

    work page 2025

  3. [3]

    L. S. Wong, J. Mao, P. Sharma, Z. S. Siegel, J. Feng, N. Korneev, J. B. Tenenbaum, and J. Andreas. Learning adaptive planning representations with natural language guidance. InThe International Conference on Learning Representations (ICLR), 2023

    work page 2023

  4. [4]

    Shridhar, J

    M. Shridhar, J. Thomason, D. Gordon, Y . Bisk, W. Han, R. Mottaghi, L. Zettlemoyer, and D. Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In 9 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740–10749, 2020

    work page 2020

  5. [5]

    RTFM: Generalising to Novel Environment Dynamics via Reading

    V . Zhong, T. RocktÃd’schel, and E. Grefenstette. Rtfm: Generalising to novel environment dynamics via reading.arXiv, 2021. URLhttps://arxiv.org/abs/1910.08210

    work page Pith review arXiv 2021

  6. [6]

    Carroll, R

    M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the utility of learning about humans for human-ai coordination.Advances in neural information processing systems, 32, 2019

    work page 2019

  7. [7]

    B. Liu, Y . Jiang, X. Zhang, Q. Liu, S. Zhang, J. Biswas, and P. Stone. LLM+P: Empowering large language models with optimal planning proficiency.ArXiv preprint, 2023

    work page 2023

  8. [8]

    Zhang, P

    L. Zhang, P. Jansen, T. Zhang, P. Clark, C. Callison-Burch, and N. Tandon. PDDLEGO: Iterative planning in textual environments. In D. Bollegala and V . Shwartz, editors,Proceedings of the 13th Joint Conference on Lexical and Computational Semantics (*SEM 2024), pages 212–221, Mexico City, Mexico, June 2024. Association for Computational Linguistics. doi: ...

    doi:10.18653/v1/2024.starsem-1.17 2024

Show all 32 references
  1. [9]

    X. Lin, Y . Wu, H. Yang, Y . Zhang, Y . Zhang, and J. Ji. CLMASP: Coupling large language models with answer set programming for robotic task planning.ArXiv, 2024

    2024

  2. [10]

    Silver, S

    T. Silver, S. Dan, K. Srinivas, J. B. Tenenbaum, L. P. Kaelbling, and M. Katz. Generalized planning in pddl domains with pretrained large language models. InAAAI Conference on Artificial Intelligence, 2023

    2023

  3. [11]

    Arora, H

    A. Arora, H. Fiorino, D. Pellier, M. M. Â. Etivier, and S. Pesty. A Review of Learning Planning Action Models.Knowledge Engineering Review, 33, 2018. doi:10.1017/S0269888918000188

    2018 doi

  4. [12]

    L. Guan, K. Valmeekam, S. Sreedharan, and S. Kambhampati. Leveraging pre-trained large language models to construct and utilize world models for model-based task planning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https: //openreview.net/...

    2023

  5. [13]

    Oswald, K

    J. Oswald, K. Srinivas, H. Kokel, J. Lee, M. Katz, and S. Sohrabi. Large language models as planning domain generators. InProceedings of the International Conference on Automated Planning and Scheduling, volume 34, pages 423–431, 2024

    2024

  6. [14]

    M. Han, Y . Zhu, S.-C. Zhu, Y . N. Wu, and Y . Zhu. Interpret: Interactive predicate learning from language feedback for generalizable task planning. InRobotics: Science and Systems (RSS), 2024

    2024

  7. [15]

    R. E. Fikes and N. J. Nilsson. Strips: A new approach to the application of theorem proving to problem solving.Artificial Intelligence, 1971

    1971

  8. [16]

    Carbonell, O

    J. Carbonell, O. Etzioni, Y . Gil, R. Joseph, C. Knoblock, S. Minton, and M. Veloso. Prodigy: An integrated architecture for planning and learning.SIGART Bull., 1991

    1991

  9. [17]

    Nau, T.-C

    D. Nau, T.-C. Au, O. Ilghami, U. Kuter, J. W. Murdock, D. Wu, and F. Yaman. Shop2: An htn planning system.J. Artif. Intell. Res. (JAIR), 2003

    2003

  10. [18]

    Brewka, T

    G. Brewka, T. Eiter, and M. Truszczy ´nski. Answer set programming at a glance.Commun. ACM, 2011

    2011

  11. [19]

    Lifschitz

    V . Lifschitz. Answer set programming and plan generation.Artificial Intelligence, 2002

    2002

  12. [20]

    G. Chen, L. Yang, R. Jia, Z. Hu, Y . Chen, W. Zhang, W. Wang, and J. Pan. Language-augmented symbolic planner for open-world task planning. InRobotics: Science and Systems Conference (RSS), 2024. 10

    2024

  13. [21]

    S. Y . Min, D. S. Chaplot, P. Ravikumar, Y . Bisk, and R. Salakhutdinov. Film: Following instructions in language with modular methods.arXiv, 2021

    2021

  14. [22]

    Zhang*, V

    T. Zhang*, V . Kishore*, F. Wu*, K. Q. Weinberger, and Y . Artzi. Bertscore: Evaluating text generation with bert. InInternational Conference on Learning Representations, 2020

    2020

  15. [23]

    M. Helmert. The fast downward planning system.J. Artif. Int. Res., 26(1):191â ˘A¸ S246, July

  16. [24]

    Introducing gpt-4o

    OpenAI. Introducing gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. URL https: //openai.com/index/hello-gpt-4o/

    2024

  17. [25]

    Introducing claude 3.7 sonnet

    Anthropic. Introducing claude 3.7 sonnet. https://www.anthropic.com/news/claude-3-7-sonnet,

  18. [26]

    Y . Chen, W. Cui, Y . Chen, M. Tan, X. Zhang, D. Zhao, and H. Wang. Robogpt: an intelligent agent of making embodied long-term decisions for daily instruction tasks.arXiv preprint arXiv:2311.15649, 2023

    2023 arXiv

  19. [27]

    Q. Team. Qwen2.5: A party of foundation models, September 2024. URL https://qwenlm. github.io/blog/qwen2.5/

    2024

  20. [28]

    Seipp, Á

    J. Seipp, Á. Torralba, and J. Hoffmann. PDDL generators, 2022

    2022

  21. [29]

    Singh, A

    I. Singh, A. Goyal, S. Birchfield, D. Fox, A. Garg, and V . Blukis. Og-vla: 3d-aware vision language action model via orthographic image generation.arXiv preprint, 2025

    2025

  22. [30]

    Y . Zhu, A. Joshi, P. Stone, and Y . Zhu. Viola: Imitation learning for vision-based manipulation with object proposal priors.arXiv preprint arXiv:2210.11339, 2022. doi:10.48550/arXiv.2210. 11339

    2022 doi arXiv

  23. [31]

    rock_paper_scissors

    P. Smirnov, F. Joublin, A. Ceravola, and M. Gienger. Generating consistent pddl domains with large language models.ArXiv preprint, 2024. 11 Appendix A Robot BlocksWorld Tasks 12 A.1 Task description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 A.2...

    2024

  24. [2025]

    URLhttps://www.anthropic.com/news/claude-3-7-sonnet

Pith tools

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