Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Language-Enhanced Mobile Manipulation for Efficient Object Search in Indoor Environments

T0 review · 4 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read This paper claims that a hierarchical, LLM-guided search lets a mobile manipulator find a target object while inspecting only about a fifth of the scene's rooms, carriers, and item surfaces.

desk verdict The LLM hierarchy idea is reasonable, but the headline efficiency claim is not supported because GODHS gets a free exhaustive pre-map while the baselines don't. read the letter →

arxiv 2508.20899 v1 pith:XDH2HPV2 submitted 2025-08-28 cs.RO

classification cs.RO
keywords mobilemanipulationobjectsearchlargelanguagemodelshierarchicalnavigationsemanticreasoningmotionplanningposesortingIsaacSimsimulation
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 is trying to show that a mobile manipulation robot can find a target object in an unfamiliar indoor scene much faster by letting a large language model order the search: which room to enter, which piece of furniture or device to inspect, and which part of it to look at first. This is the GODHS framework, a five-level hierarchy from scene to room to carrier to feature to item. In simulations of a seven-zone flat, the LLM-guided system's overall search rate was about 21–23% of the environment versus 60.5% for full coverage and 51.8% for random walking, meaning it inspected far fewer rooms, carriers, and items before finding the target. The paper also contributes a heuristic motion planner that orders chassis and end-effector poses so the robot moves around a carrier smoothly, cutting execution time to 66% of the unsorted baseline. If true, this makes language commonsense a practical front-end for object search without exhaustive mapping.

What carries the argument

The five-level hierarchy (scene → room → carrier → feature → item) is the load-bearing structure; each level applies an LLM ranking equation to order a small candidate list, and a structured prompt with output constraints keeps the answer usable. Supporting machinery includes geometric feature extraction from carrier point clouds (top, sides, bottom, inside), greedy pose selection for visual coverage, inverse-kinematics validation, lexicographical end-effector pose ordering, and centroid-aligned polar-angle chassis sorting so the base circles the carrier clockwise rather than jumping across it.

What would settle it

Place the target deliberately in the lowest-ranked room, carrier, and feature (e.g., an orange on a study bookshelf rather than in the fridge) and run the same 81-experiment protocol; if the overall search rate rises to or above the 51–60% random and coverage baselines, the claimed efficiency gain is not robust to prior-unfriendly placements.

Watch

Extended reading notes

Core claim

The central claim is that semantic commonsense, injected through an LLM at each level of a bounded search hierarchy, can substitute for exhaustive geometric exploration. The robot first maps the flat with LiDAR and semantic segmentation; the LLM infers room categories, ranks rooms by likelihood of the target, then ranks candidate carrier objects within a room, then ranks spatial features of each carrier. Only after these ordered lists are fixed does the robot physically move and inspect, so each level prunes the search space. The hierarchical ranking is made reliable by structured prompts that constrain outputs to predefined sets and machine-readable formats, plus a clean-process-correct cyc

Load-bearing premise

The search savings depend on the LLM's rankings of rooms, carriers, and features being better than random in the actual scene; if the commonsense priors point to the wrong places or the segmenter mislabels objects, the hierarchy only adds overhead.

Editorial extensions

If this is right

  • In an unfamiliar indoor scene, telling the robot what the target is can cut required exploration by more than half compared to coverage or random baselines, even with a small 7B LLM.
  • The LLM's search order can be dynamically re-ranked as new rooms and carriers are observed, so the system degrades gracefully when a high-priority location turns out empty.
  • The pose-sorting planner can be used independently of the LLM layer: chassis polar-angle sorting plus end-effector lexicographical sorting together reduce the execution-time ratio to 0.66 and improve both path-length ratios.
  • Structured prompt constraints make LLM outputs machine-parseable, closing the gap between free-text commonsense and robot commands without fine-tuning.
  • Search cost decomposes into room, carrier, and item rates, so the approach can be measured and tuned at each level separately.

Reading between the lines

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

  • The paper's single simulated flat leaves room for a stress test: place the target in the semantically least likely spot, e.g., an orange on a study bookshelf rather than in the fridge, and the OSR gap over coverage search should shrink or reverse; that would delimit when LLM priors help.
  • The same room→carrier→feature decomposition could be driven by a vision-language model or a hybrid LLM-VLM, which might remove the separate semantic-segmentation dependency and handle open-vocabulary target descriptions.
  • The OSR weights (0.2, 0.3, 0.5) are an arbitrary choice; reweighting toward room cost or item cost could change the apparent advantage, so comparisons should report the three rates separately, as the paper does.
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 / 7 minor

Summary. The paper proposes GODHS, a language-enhanced hierarchical search framework for mobile manipulators. The method decomposes object search into five levels (scene, room, carrier, feature, item) and uses an LLM to rank search priorities at the room, carrier, and feature levels. A heuristic motion planner generates chassis and end-effector poses and sorts them via polar-angle and lexicographical sorting to reduce redundant motion. The system is evaluated in Isaac Sim on a flat scene, with the target an orange inside a fridge, using two LLMs (GPT-4o and Qwen2.5-7B) against coverage and random-walk baselines. The paper reports significantly lower Room/Carrier/Item Search Rates and a lower OSR for GODHS and concludes that the method achieves higher search efficiency.

Significance. The core idea is timely and potentially useful: using LLM commonsense priors to prune a hierarchical search space could reduce object search effort in unknown indoor environments. The paper's strengths are its clear five-level hierarchy, the use of structured prompts to constrain LLM outputs, and a concrete pose-generation and sorting scheme that shows an internal ablation improvement in path length and execution time. However, the central empirical claim is not supported as presented. The evaluation gives GODHS a free pre-mapping phase that visits every room before the measured search begins, while the coverage and random-walk baselines must explore from scratch. The paper also omits success rates, variance, significance tests, and sufficient baseline details. If the pre-mapping confound is removed or properly controlled and the missing statistics are supplied, the approach could be a useful contribution; as written, the efficiency comparison is uninterpretable.

major comments (4)
  1. [Algorithm 1, Lines 2–8; Eq. (8); Table I] The main efficiency comparison is confounded by the pre-mapping phase. Lines 2–8 of Algorithm 1 state that the robot first visits and maps all accessible rooms, obtaining a complete scene map and semantic labels, before the prioritized phase begins. The OSR metric in Eq. (8) appears to count only rooms/carriers/items searched in the later prioritized phase (Lines 9–28), so the exhaustive mapping traversal is excluded from the cost. The coverage and random-walk baselines are not given a pre-built map; they must physically explore the environment to find the target. Thus Table I's large gap (21–23% vs. 60%/52%) conflates the value of the LLM prior with the value of free access to a complete map and semantic segmentation. To support the central efficiency claim, the paper should report total task time or distance, or compare against baselines that also perform the same pre-mapping phase and
  2. [Table I, Sec. IV-B] The paper reports means over 81 runs but gives no error bars, standard deviations, or significance tests. It also does not report the success rate; failures are acknowledged in Sec. IV-B, but it is unclear whether the Table I numbers are conditional on success or how failed runs enter the metric. A method that finds the target only in a few easy cases could have low search rates on those successes while being inefficient overall. The authors should provide per-run distributions, success rates, and appropriate statistical tests for the GODHS-vs-baseline comparison.
  3. [Sec. III-B, Algorithm 1, Fig. 5] The 'inside' feature of an enclosed fridge requires opening the fridge door, and Figure 5's caption says the robot 'opens the door' to find the orange. However, no manipulation primitive, door-opening controller, or physical interaction is described anywhere in the method; the motion planner only generates visual-inspection poses for the chassis and end-effector. If door-opening is required, it must be explicitly modeled and accounted for in the evaluation. Without this, the feasibility of the central demonstration is incomplete.
  4. [Sec. IV-B, Table I] The coverage and random-walk baselines are described too briefly to be reproducible. The paper does not specify how coverage search is defined at each hierarchy level (rooms/carriers/items), how random walk is parameterized, what termination criteria are used, or whether the baselines use semantic segmentation. Without this detail, the numerical gap in Table I cannot be independently verified or interpreted.
minor comments (7)
  1. [Eq. (8)] The OSR weights w1=0.2, w2=0.3, w3=0.5 are chosen by hand with no sensitivity analysis. Since all three component rates improve for both LLMs, the qualitative conclusion is robust to these weights, but the justification and sensitivity should be reported.
  2. [Eq. (6)] The bottom-area plane height z_F0 is not defined or motivated. Please state how it is chosen and whether it varies across carriers.
  3. [Sec. III-B] The angle computation for field-of-view coverage uses arccos with dot products mixing x/y/z components; the notation and axes are unclear and should be corrected or rewritten in a standard form.
  4. [Algorithm 1] In Lines 3–8, rooms are sorted inside the mapping loop. If the map is being completed as new rooms are entered, the room ordering should be recomputed after the loop; otherwise the final prioritized order may not reflect the complete map.
  5. [Sec. III-B] The term 'dictionary mapping' is introduced but not defined. It is unclear whether this refers to the P_EE^CH dictionary or to the room-to-map/carrier-to-map correspondences.
  6. [References] Reference [25] cites Dijkstra for the greedy pose-selection algorithm; the connection is not obvious. Please cite the appropriate greedy set-cover or viewpoint-selection method.
  7. [Fig. 5] The figure is dense and the text in the subpanels is very small; it would benefit from higher resolution or a cleaner layout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LLM priors are external, OSR is an evaluation metric, and self-citations are background.

full rationale

GODHS's central claim is that LLM-ordered hierarchical search inspects fewer rooms, carriers, and features before finding a target. The ordering is produced by pre-trained LLMs (GPT-4o, Qwen2.5-7B) through Eqs. (1)-(3), and no parameters are fitted to the Table I outcomes. OSR (Eq. 8) is a measurement of the resulting search order, not a quantity that is defined to equal the claimed efficiency; the weights (0.2/0.3/0.5) are arbitrary but do not enter into the generation of the search orders. No uniqueness theorem or prior result by the same authors is invoked to force the method. The self-citations in Related Work (e.g., [2], [6], [7], [9], [18]) support standard motion-planning background and do not carry the search-efficiency conclusion. The pre-mapping phase in Algorithm 1 (Lines 3-8) and the baselines' not receiving a pre-built map could bias the comparison, but that is an experimental confound, not a circular derivation: the conclusion is not logically identical to its assumptions. Therefore, under the required evidence standard, no circular step can be exhibited.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim leans on standard geometry and an LLM's semantic priors. The only numeric free parameters are the arbitrary OSR weights and the unspecified bottom-plane height z_F0. No new physical entities are introduced.

free parameters (2)
  • OSR weights w1, w2, w3 = 0.2, 0.3, 0.5
    Hand-chosen weights in Eq. 8 used to aggregate room, carrier, and item search rates into the headline Overall Search Rate; different weights could change the comparison.
  • Bottom-area plane height z_F0 = not specified
    Eq. 6 defines the bottom feature as a plane at fixed height z_F0, but no value or estimation procedure is given; it affects which points are searched.
assumptions (4)
  • domain assumption LLM commonsense rankings transfer from language training to the robot's semantic observations
    Eqs. 1-3 and the prompt design in Sec. III-A assume LLM priors like 'orange is likely in a fridge' lead to correct search order in the test scene.
  • domain assumption Semantic segmentation and object detection correctly identify all rooms, carriers, and target items
    Algorithm 1 lines 7, 11, and 27 rely on SemSeg of camera data; failures in segmentation would break the hierarchy.
  • domain assumption The scene map can be completed by visiting all accessible rooms
    Algorithm 1 lines 3-8 loop until the scene map is complete, assuming every relevant room is reachable and mappable.
  • domain assumption The five-level hierarchy room-carrier-feature is the correct inductive bias for object search
    The method and evaluation assume decomposing search this way is more efficient than flat search; no ablation against alternative hierarchies is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Language-Enhanced Mobile Manipulation for Efficient Object Search in Indoor Environments." pith.science (2026). https://pith.science/paper/XDH2HPV2

@misc{pith2026250820899,
  author       = {Pith},
  title        = {Pith review of: Language-Enhanced Mobile Manipulation for Efficient Object Search in Indoor Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XDH2HPV2}},
  note         = {Machine review of arXiv:2508.20899}
}
read the original abstract

Enabling robots to efficiently search for and identify objects in complex, unstructured environments is critical for diverse applications ranging from household assistance to industrial automation. However, traditional scene representations typically capture only static semantics and lack interpretable contextual reasoning, limiting their ability to guide object search in completely unfamiliar settings. To address this challenge, we propose a language-enhanced hierarchical navigation framework that tightly integrates semantic perception and spatial reasoning. Our method, Goal-Oriented Dynamically Heuristic-Guided Hierarchical Search (GODHS), leverages large language models (LLMs) to infer scene semantics and guide the search process through a multi-level decision hierarchy. Reliability in reasoning is achieved through the use of structured prompts and logical constraints applied at each stage of the hierarchy. For the specific challenges of mobile manipulation, we introduce a heuristic-based motion planner that combines polar angle sorting with distance prioritization to efficiently generate exploration paths. Comprehensive evaluations in Isaac Sim demonstrate the feasibility of our framework, showing that GODHS can locate target objects with higher search efficiency compared to conventional, non-semantic search strategies. Website and Video are available at: https://drapandiger.github.io/GODHS

Figures

Figures reproduced from arXiv: 2508.20899 by the authors.

Figure 1
Figure 1. The GODHS framework divides each scene into five strict levels: Scene, Room, Carrier, Feature, and Item. For instance, after entering the flat, the robot uses mapping and semantic segmentation to determine and prioritize the types of rooms through a LLM. For the first room kitchen, the LLM classifies and prioritizes carriers such as furniture or devices. For the first carrier fridge, the LLM identifies and prioritiz… view at source ↗
Figure 2
Figure 2. Complete Architecture of the GODHS System: First, the user provides the target object through natural language input, which is then processed by a LLM to extract semantic intent. The robot captures geometric data using LiDAR and RGB images via cameras, generating a topological map with room names through the language model. Within the known map, the LLM-driven prioritization ranks rooms based on the likelihood of co… view at source ↗
Figure 3
Figure 3. LLM Reasoning Process (Left): A multi-stage process involving data cleaning, processing with optional knowledge aug￾mentation, and correcting with optional feedback ensures reliable output. Prompt Design (Right): An example of a structured prompt used to ground the LLM’s reasoning for a specific task. B. Heuristic-Based Pose Generation and Sorting This subsection details our heuristic-based methodology for computing… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Leftmost: Determine EE poses via carrier geometry analysis. Second Left: Generate CH poses through greedy EE pose exploration. Second Right: Verify CH-EE pairs with inverse kinematics validation. Rightmost: Prioritize CH poses by Polar Angle Sorting and EE poses by Lex…
Figure 5
Figure 5. Figure 5: Feasibility Example: Taking a Qwen2.5-7B-powered search as an example, the target is an orange. After an initial exploration to map the scene (lower left), the LLM guides the robot to the living room first. It inspects the coffee table’s top and bottom surfaces without…

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. Relational Semantic Reasoning on 3D Scene Graphs for Open World Interactive Object Search

    cs.RO 2026-03 accept novelty 6.0 of 10

    SCOUT matches LLM planners on open-world interactive object search by scoring 3D scene-graph nodes with lightweight models distilled from LLM relational priors, at far lower compute cost.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    The spatial semantic hierarchy,

    B. Kuipers, “The spatial semantic hierarchy,” Artificial intelligence , vol. 119, no. 1-2, pp. 191–233, 2000

  2. [2]

    Motion planning for robotics: A review for sampling-based planners,

    L. Zhang, K. Cai, Z. Sun, Z. Bing, C. Wang, L. Figueredo, S. Had- dadin, and A. Knoll, “Motion planning for robotics: A review for sampling-based planners,” Biomimetic Intelligence and Robotics , vol. 5, no. 1, p. 100207, 2025

  3. [3]

    Using intermediate objects to improve the efficiency of visual search,

    L. E. Wixson and D. H. Ballard, “Using intermediate objects to improve the efficiency of visual search,”Int. J. Comput. Vision, vol. 12, no. 2–3, p. 209–230, Apr. 1994

  4. [4]

    Estimated informed anytime search for sampling-based planning via adaptive sampler,

    L. Zhang, K. Cai, Y . Zhang, Z. Bing, C. Wang, F. Wu, S. Haddadin, and A. Knoll, “Estimated informed anytime search for sampling-based planning via adaptive sampler,” IEEE Transactions on Automation Science and Engineering , vol. 22, pp. 18 580–18 593, 2025

  5. [5]

    Autonomous search for mines,

    E. Gelenbe and Y . Cao, “Autonomous search for mines,” European Journal of Operational Research , vol. 108, no. 2, pp. 319–333, 1998

  6. [6]

    APT*: Asymptotically optimal motion planning via adaptively prolated elliptical r-nearest neighbors,

    L. Zhang, S. Wang, K. Cai, Z. Bing, F. Wu, C. Wang, S. Haddadin, and A. Knoll, “APT*: Asymptotically optimal motion planning via adaptively prolated elliptical r-nearest neighbors,” IEEE Robotics and Automation Letters, vol. 10, no. 10, pp. 10 242–10 249, 2025

  7. [7]

    Tree- based grafting approach for bidirectional motion planning with local subsets optimization,

    L. Zhang, Y . Ling, Z. Bing, F. Wu, S. Haddadin, and A. Knoll, “Tree- based grafting approach for bidirectional motion planning with local subsets optimization,” IEEE Robotics and Automation Letters, vol. 10, no. 6, pp. 5815–5822, 2025

  8. [8]

    Demonstration to adaptation: A user-guided framework for sequential and real-time planning,

    K. Cai, R. Laha, Y . Gong, L. Chen, L. Zhang, L. F. Figueredo, and S. Haddadin, “Demonstration to adaptation: A user-guided framework for sequential and real-time planning,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 9871– 9878

Show all 29 references
  1. [9]

    Genetic informed trees (GIT*): Path planning via reinforced genetic programming heuristics,

    L. Zhang, K. Cai, Z. Bing, C. Wang, and A. Knoll, “Genetic informed trees (GIT*): Path planning via reinforced genetic programming heuristics,” Biomimetic Intelligence and Robotics , vol. 5, no. 3, p. 100237, 2025

  2. [10]

    Act to see and see to act: Pomdp planning for objects search in clutter,

    J. K. Li, D. Hsu, and W. S. Lee, “Act to see and see to act: Pomdp planning for objects search in clutter,” pp. 5701–5707, 2016

  3. [11]

    Language- guided semantic mapping and mobile manipulation in partially ob- servable environments,

    S. Patki, E. Fahnestock, T. M. Howard, and M. R. Walter, “Language- guided semantic mapping and mobile manipulation in partially ob- servable environments,” 2019

  4. [12]

    Remote object navigation for service robots using hierarchical knowledge graph in human-centered environments,

    Y . Li, Y . Ma, X. Huo, and X. Wu, “Remote object navigation for service robots using hierarchical knowledge graph in human-centered environments,”Intelligent Service Robotics, vol. 15, pp. 1–15, 06 2022

  5. [13]

    Visual language maps for robot navigation,

    C. Huang, O. Mees, A. Zeng, and W. Burgard, “Visual language maps for robot navigation,” 2023

  6. [14]

    Goat: Go to any thing,

    M. Chang, T. Gervet, M. Khanna, S. Yenamandra, D. Shah, S. Y . Min, K. Shah, C. Paxton, S. Gupta, D. Batra, R. Mottaghi, J. Malik, and D. S. Chaplot, “Goat: Go to any thing,” 2023

  7. [15]

    Openin: Open-vocabulary instance-oriented navigation in dynamic domestic environments,

    Y . Tang, M. Wang, Y . Deng, Z. Zheng, J. Deng, and Y . Yue, “Openin: Open-vocabulary instance-oriented navigation in dynamic domestic environments,” 2025

  8. [16]

    Llm-planner: Few-shot grounded planning for embodied agents with large language models,

    C. H. Song, J. Wu, C. Washington, B. M. Sadler, W.-L. Chao, and Y . Su, “Llm-planner: Few-shot grounded planning for embodied agents with large language models,” 2023

  9. [17]

    Language-grounded dynamic scene graphs for interactive object search with mobile manipulation,

    D. Honerkamp, M. B ¨uchner, F. Despinoy, T. Welschehold, and A. Val- ada, “Language-grounded dynamic scene graphs for interactive object search with mobile manipulation,” IEEE Robotics and Automation Letters, vol. 9, no. 10, p. 8298–8305, Oct. 2024

  10. [18]

    Elliptical k-nearest neighbors - path optimization via coulomb’s law and invalid vertices in c-space obstacles,

    L. Zhang, Z. Bing, Y . Zhang, K. Cai, L. Chen, F. Wu, S. Haddadin, and A. Knoll, “Elliptical k-nearest neighbors - path optimization via coulomb’s law and invalid vertices in c-space obstacles,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)...

  11. [19]

    POMP: pomcp-based online motion plan- ning for active visual search in indoor environments,

    Y . Wang, F. Giuliari, R. Berra, A. Castellini, A. D. Bue, A. Farinelli, M. Cristani, and F. Setti, “POMP: pomcp-based online motion plan- ning for active visual search in indoor environments,” CoRR, vol. abs/2009.08140, 2020

  12. [20]

    A data structure for dynamic trees,

    D. D. Sleator and R. Endre Tarjan, “A data structure for dynamic trees,” Journal of Computer and System Sciences , vol. 26, no. 3, pp. 362–391, 1983

  13. [21]

    Active visual object search in unknown environments using uncertain seman- tics,

    A. Aydemir, A. Pronobis, M. G ¨obelbecker, and P. Jensfelt, “Active visual object search in unknown environments using uncertain seman- tics,” IEEE Transactions on Robotics , vol. 29, no. 4, pp. 986–1002, 2013

  14. [22]

    Training language models to follow instruc- tions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe, “Training language models to follow instruc- tio...

  15. [23]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” 2023

  16. [24]

    Self-refine: Iterative refinement with self-feedback,

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yang, S. Gupta, B. P. Majumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P. Clark, “Self-refine: Iterative refinement with self-feedback,” 2023

  17. [25]

    A note on two problems in connexion with graphs

    E. DIJKSTRA, “A note on two problems in connexion with graphs.” pp. 269–271, 1959

  18. [26]

    Inverse kinematic solutions with singularity robustness for robot manipulator control,

    Y . Nakamura and H. Hanafusa, “Inverse kinematic solutions with singularity robustness for robot manipulator control,” 1986

  19. [27]

    An efficient algorithm for determining the convex hull of a finite planar set,

    R. L. Graham, “An efficient algorithm for determining the convex hull of a finite planar set,” Inf. Process. Lett. , vol. 1, pp. 132–133, 1972

  20. [28]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli et al., “On the opportunities and risks of foundation models,” 2022

  21. [29]

    Curiosity- based robot navigation under uncertainty in crowded environments,

    K. Cai, W. Chen, C. Wang, H. Zhang, and M. Q.-H. Meng, “Curiosity- based robot navigation under uncertainty in crowded environments,” IEEE Robotics and Automation Letters , vol. 8, no. 2, pp. 800–807, 2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.