Pith. sign in

REVIEW 4 major objections 5 minor 20 references

The paper claims that a living plant and a mobile robot, joined by asynchronous LLM modules that converse in natural language, behave as one hybrid agent, with the plant's physiological needs supplying the goals.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A plant and a mobile robot are coupled through a modular network of LLM agents that translate sensor data into natural-language messages and then into motor commands.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A working plant-robot hybrid driven by modular LLM agents is genuinely new as a demo, but the paper's claim that the plant's normativity 'installs' agency is not backed by any control condition. the 4 major comments →

arxiv 2509.05338 v1 pith:CPRY6AZP submitted 2025-09-01 cs.RO cs.AI

Plantbot: Integrating Plant and Robot through LLM Modular Agent Networks

Atsushi Masumori , Norihiro Maruyama , Itsuki Doi , johnsmith , Hiroki Sato , Takashi Ikegami This is my paper
classification cs.RO cs.AI
keywords plant-robot hybridLLM modular agentsnatural-language protocolnormativitydistributed agencymobile robotartificial life
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 introduces Plantbot, a working prototype that connects a living plant and its soil to a mobile robot through a network of asynchronous large language model modules that pass messages to one another in natural language. The central claim is that natural language can act as a universal protocol layer, binding a biological system and a machine into a single hybrid agent: the sensor module speaks for the soil, the chat module interprets, and the action modules move the robot. The plant's real physiological needs—chiefly thirst—give this loop a direction, which the authors argue installs normativity into the sensorimotor loop and grounds agency in the living substrate. If the claim holds, LLM-based language becomes a general-purpose glue for building embodied hybrid lifeforms from heterogeneous components.

Core claim

On its own terms, the paper claims that a distributed network of five asynchronous LLM modules—sensing, vision, central chat, and two action layers—can transform a plant's physiological state into a mobile robot's behavior, and that the resulting system constitutes a unified embodied agent rather than a set of disconnected parts. The demonstration is an installation: soil moisture, temperature, pH, nutrient levels, and camera frames are converted into natural-language messages; the Chat Agent combines these with human dialogue; Action Agents then issue stop/move decisions and concrete motor commands. The authors interpret their logged utterances, word clouds, and embedding clusters as eviden

What carries the argument

The central object is an asynchronous modular network of five LLM agents—Vision, Sensor, Chat, Action 1, and Action 2—that communicate exclusively through natural-language messages over OSC. Natural language acts as the universal protocol layer: heterogeneous multimodal data (soil readings, camera frames, human speech) are translated into linguistic messages, the Chat Agent integrates them into high-level directives, and the Action Agents convert those directives into motor commands. The plant's physiological norm—its need for water—supplies the direction for the whole loop, and the architecture's claim is that this norm is carried through language and installed in the robot's sensorimotor c

Load-bearing premise

The central claim depends on the plant's measured state actually steering the language-model decisions, rather than the system's stop-and-move behavior being driven mostly by random variation, conversation context, or the way the prompts happen to be written.

What would settle it

Replace the live soil sensor with a pre-recorded or randomized stream of identical-looking readings while keeping the same prompts and conversation history; if the robot's stop/move behavior and its linguistic triggers remain statistically unchanged, then the plant's state is not causally responsible for the behavior the paper assigns to it.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Adding a new sensor or actuator to Plantbot requires only a textual description of its role; the language protocol does the integration work without redesigning low-level interfaces.
  • Dry soil can cause the robot to move in search of water through an autonomous message cascade, without any explicit human instruction.
  • Because each module is semantically specialized, the chain from sensor reading to motor command is human-legible and can be audited through the agents' own text.
  • The architecture can be scaled into collectives where plants, robots, sensors, and humans all use the same natural-language layer to coordinate.
  • Future versions could let modules rewire their own communication topology through meta-level dialogue, enabling self-organizing changes without retraining the underlying models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The authors' word-cloud and transcript evidence is correlational; a stronger causal test would compare live-plant trials against identical trials with a dead plant or a randomized sensor stream, holding prompts and conversation history fixed.
  • If natural language is truly a universal protocol layer, the same architecture should work with other biological signal sources—bioelectric potentials, fungal mycelia, or microbial sensors—without changing the action modules.
  • The proposed 'plastic' communication topology is testable: letting agents negotiate their own message routing should produce network structures that differ across environments, and those differences could be linked to behavioral performance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents Plantbot, a mobile robot with a living plant, in which five asynchronously operating LLM modules (Vision, Sensor, Chat, Action 1, Action 2) exchange natural-language messages. Sensor readings and visual context are converted into linguistic statements, a central Chat Agent integrates them with human dialogue, and two Action Agents produce stop/move commands for a tracked base. The system was deployed in two public exhibitions, and logs were analyzed with UMAP, word clouds, state-transition counts, and three illustrative transcripts. The authors argue that natural language functions as a universal protocol layer between biological and artificial subsystems, and that the plant's physiological needs install normativity, making Plantbot a single unified hybrid agent.

Significance. The architectural idea is timely and the public deployments represent a real engineering effort. If the central claim were established, the paper would contribute a useful concept to ALife/HRI: LLM modules as a shared linguistic protocol for heterogeneous embodied components. The paper is also transparent about its prompt-based design and provides concrete transcripts, which is a strength. However, the empirical sections do not test the causal claim that the plant's physiological state drives the robot's actions; they show only that role-specific prompts yield role-specific language and that action states correlate with chat messages. The paper's value is therefore best assessed as a system demonstration, not as a test of agency or normativity.

major comments (4)
  1. [Experiments (Figs. 4-7)] There is no baseline or control condition. The UMAP clustering by agent role and the word-cloud differences are expected consequences of the distinct role prompts described in the 'LLM Modules Network' section: the Sensor Agent is told to produce emotive soil reports, the Chat Agent is told it is Plantbot, and the Action Agents are told to issue move/stop decisions. These analyses therefore do not distinguish plant-driven behavior from prompt-driven behavior. To support the claim that the Action module is 'context-sensitive' and that dry soil leads to movement, the authors need a comparison with a dead plant, a fake/simulated sensor stream, or a reversed-message condition, with action selection as the dependent variable.
  2. [Experiments (Examples 1-3)] The three transcripts are hand-picked and cannot establish a causal cascade. In Example 1, the Sensor message 'The soil is dry' and the Chat response 'Could you water me?' closely follow the Sensor Agent's explicit prompt instruction to convert soil data into emotive statements such as 'I am thirsty.' The later Action command to move toward moisture is thus partly a product of prompt engineering. The authors should report how often such sensor-to-action transitions occur across all logs, and should test whether actual sensor values predict actions after conditioning on the prompt-generated language.
  3. [Figure 6 and surrounding text] The text states that stop states occurred 'significantly more often' and that state durations differ 'markedly,' but no statistical test, sample size, variance, or null model is provided. Without a baseline of random action selection or a shuffled-message comparison, the stop/move counts and the word clouds in Figure 7 do not demonstrate that decisions are based on 'semantically meaningful cues.' Please provide transition probabilities conditioned on sensor/chat message categories, effect sizes, and appropriate statistical comparisons.
  4. [Abstract and Discussion] The central claim that the plant's normativity is 'installed' and that behavior 'emerges not from direct human instruction, but from the structure itself' conflates the plant's biological needs with the LLM's instruction-following. The Sensor Agent is explicitly prompted to speak as the plant, and the Chat Agent is explicitly prompted to believe it is Plantbot. This may be an interesting way to implement a norm in a hybrid system, but it does not support the conclusion that the plant's physiology is the source of the agent's behavior. Either soften the claim to 'prompt-implemented normativity' or provide an intervention showing that plant-derived signals (e.g., soil moisture thresholds) have a measurable effect on actions independent of the prompt text.
minor comments (5)
  1. [Figure 4/5] Please report the number of utterances used for embedding, UMAP hyperparameters, random seed, and whether the clustering is stable across trials. Currently the figure legends are qualitative.
  2. [Discussion] 'wailing plant' appears to be a typo; it should likely be 'wilting plant.'
  3. [System section] The notation is inconsistent: 'Action agent 1/2' and 'Action Agent 1/2' are used interchangeably. Use a single capitalization and abbreviation style.
  4. [Experiments footnote] The note that original Japanese outputs were translated to English is useful, but it should specify which analyses were performed on translated text and how translation may affect embeddings and word clouds.
  5. [Discussion] The term 'hybrid lifeform' is used as a conclusion-bearing concept. Provide an operational definition, e.g., a closed sensorimotor loop across biological and artificial components, so the claim is falsifiable.

Circularity Check

2 steps flagged

The plant-driven 'emergent agency' claim reduces by construction to the authors' own prompt definitions: the dry-soil-to-move chain is installed verbatim in the Sensor, Chat, and Action prompts, then reported as evidence of plant-installed normativity.

specific steps
  1. self definitional [System/LLM Modules Network (Section 2.2) and Discussion]
    "Sensor Agent (GPT-3.5 Turbo). Converts quantitative soil data into short, emotive statements such as “I am thirsty.” ... Action Agent 1 (GPT-3.5 Turbo). Determines whether movement is necessary ... Prompt engineering ensures that Plantbot recognizes itself as a fusion of plant and robot, prompting it to behave as a single, life-like entity."

    The paper presents the dry-soil → 'I am thirsty' → 'I should move' chain as evidence that plant physiology installs agency ('behavior emerges not from direct human instruction, but from the structure itself'). But that chain is implemented by the authors' own prompts: the Sensor prompt is told to emit thirst when dry, the Chat prompt is told to treat the plant as thirsty and forward commands, and the Action prompt is told to decide to move. The experimental transcript is therefore an execution of those prompt definitions, not an independent prediction derived from the plant's biology. The central claim of plant-installed normativity reduces to the prompt engineering the authors supplied.

  2. self definitional [Experiments, Figure 5 analysis]
    "The distribution of terms indicates that each agent engages in semantically distinct linguistic behavior aligned with its assigned role."

    Each agent's prompt assigns it a functional role and instructs it to use role-specific vocabulary (Sensor about soil, Action about movement, etc.). The word-cloud analysis shows that the agents use the vocabulary their prompts prescribe. This is a restatement of the prompt design, not an emergent or independently discovered property. The observed 'semantic specialization' is true by construction.

full rationale

The paper's empirical core is a proof-of-concept demonstration of a prompt-defined pipeline: Sensor Agent -> Chat Agent -> Action Agent. The examples, word clouds, and UMAP all confirm that the modules follow their role prompts. The discussion then interprets this as 'normativity essential for agency' arising from the living plant and as behavior 'not from direct human instruction, but from the structure itself.' That interpretation is circular because the structure's behavior is a direct instantiation of the instructions the authors wrote into the prompts. The paper does not fit quantitative parameters or make a statistical prediction, so 'fitted input called prediction' is not a precise label; the more accurate pattern is self-definitional: the observed coupling between plant state and robot action is defined by the prompts. Self-citations to Maruyama et al. (2025) and Takata et al. (2024) are background and not load-bearing, so they do not raise the score further. Nevertheless, the central agency claim is substantially underdetermined by the evidence as presented, and the main claimed result—that the plant's biological norm 'installs' agency—collapses into the authors' own prompt engineering. Score 6 reflects partial but significant circularity: the system works as designed, but the design already contains the mapping that is later presented as an emergent discovery.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The system's observable behavior is almost entirely determined by two things the authors chose by hand: the prompt templates and the message cadence, plus one empirical constant (history length n=10). None of these are derived, ablated, or independently grounded. The conceptual entities 'hybrid lifeform' and 'installed normativity' are interpretations of the same data rather than independently evidenced additions.

free parameters (3)
  • History buffer length n=10 = 10
    The paper states the buffer length was 'chosen empirically to balance responsiveness and context retention' with no ablation or criterion.
  • Agent role prompts = not disclosed verbatim
    The behavioral outcomes (which words appear in word clouds, which actions follow which sensor states) are direct functions of the prompt templates chosen by the authors. The prompts are the effective fitted parameters of the system.
  • Decision thresholds / message cadence = not specified
    The paper does not specify when or how often sensor and vision modules emit messages, which determines the stop/move state statistics reported in Figure 6.
axioms (4)
  • domain assumption GPT-3.5 and GPT-4V, accessed through closed APIs at unspecified versions, behave consistently enough to reproduce the reported interactions.
    Enters in the LLM Modules Network section; the entire behavior depends on proprietary, non-pinned model versions.
  • domain assumption UMAP projection of utterance embeddings reflects meaningful semantic specialization rather than prompt template artifacts.
    Invoked in Figure 4. Since each agent is given a distinct role prompt, the clustering is expected by construction; the paper treats it as informative.
  • domain assumption The plant's physiological state is the causal driver of the robot's actions rather than the LLM's autocomplete tendencies.
    The whole agency/normativity narrative in the Discussion depends on this, but no ablation, control, or counterfactual test is presented.
  • domain assumption The transcribed interaction logs are representative and not cherry-picked.
    The three examples in the Experiments section are selected without a sampling protocol.
invented entities (2)
  • Concept of 'hybrid lifeform' as an emergent unified agent no independent evidence
    purpose: Reframe the engineered sensor-to-prompt-to-motor pipeline as a new kind of biological-artificial agency.
    No operational definition, no measurement, and no falsifiable prediction distinguishes 'hybrid lifeform' from a prompted robot reacting to a sensor.
  • Installed normativity from the plant's survival needs no independent evidence
    purpose: Ground the robot's goal-directed behavior in the plant's physiological norm of seeking water.
    The norm is described in prose and illustrated by two selected transcripts; there is no test showing the system behaves differently when the plant is dead, when soil data are fake, or when the norm is absent from the prompts.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Plantbot: Integrating Plant and Robot through LLM Modular Agent Networks." pith.science (2026). https://pith.science/paper/CPRY6AZP

@misc{pith2026250905338,
  author       = {Pith},
  title        = {Pith review of: Plantbot: Integrating Plant and Robot through LLM Modular Agent Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CPRY6AZP}},
  note         = {Machine review of arXiv:2509.05338}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce Plantbot, a hybrid lifeform that connects a living plant with a mobile robot through a network of large language model (LLM) modules. Each module - responsible for sensing, vision, dialogue, or action - operates asynchronously and communicates via natural language, enabling seamless interaction across biological and artificial domains. This architecture leverages the capacity of LLMs to serve as hybrid interfaces, where natural language functions as a universal protocol, translating multimodal data (soil moisture, temperature, visual context) into linguistic messages that coordinate system behaviors. The integrated network transforms plant states into robotic actions, installing normativity essential for agency within the sensor-motor loop. By combining biological and robotic elements through LLM-mediated communication, Plantbot behaves as an embodied, adaptive agent capable of responding autonomously to environmental conditions. This approach suggests possibilities for a new model of artificial life, where decentralized, LLM modules coordination enable novel interactions between biological and artificial systems.

Figures

Figures reproduced from arXiv: 2509.05338 by Atsushi Masumori, Hiroki Sato, Itsuki Doi, Johnsmith, Norihiro Maruyama, Takashi Ikegami.

Figure 1
Figure 1. Figure 1: Installation view of Plantbot, composed of a liv [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hardware setup of the plantbot system. The image [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: UMAP projection of agent utterance embed [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Word clouds of messages generated by each LLM [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Analysis of action-state dynamics. Left: Mean [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Word clouds of Chat Agent messages immediately [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages

  1. [1]

    Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Ho, D., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E., Ruano, R. J., Jeffrey, K., Jesmonth, S., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Lee, K.-H., Levine, S., Lu, Y., Luu, L., Parada, C., Pastor, P., Quiamba...

    Pith/arXiv arXiv 2022

  2. [2]

    Y., Willows, M., Yang, F., and Yang, G

    AL, A., Ahn, A., Becker, N., Carroll, S., Christie, N., Cortes, M., Demirci, A., Du, M., Li, F., Luo, S., Wang, P. Y., Willows, M., Yang, F., and Yang, G. R. (2024). Project sid: Many-agent simulations toward ai civilization

    work page 2024

  3. [3]

    Baars, B. J. (1988). A Cognitive Theory of Consciousness . Cambridge University Press, New York

    work page 1988

  4. [4]

    Baars, B. J. (2005). Global workspace theory of consciousness: toward a cognitive neuroscience of human experience. In Laureys, S., editor, The Boundaries of Consciousness: Neurobiology and Neuropathology , volume 150 of Progress in Brain Research , pages 45--53. Elsevier

    work page 2005

  5. [5]

    Baltieri, M., Iizuka, H., Witkowski, O., Sinapayen, L., and Suzuki, K. (2023). Hybrid life: Integrating biological, artificial, and cognitive systems. WIREs Cognitive Science , 14(6):e1662

    work page 2023

  6. [6]

    E., Paolo, E

    Barandiaran, X. E., Paolo, E. D., and Rohde, M. (2009). Defining agency: Individuality, normativity, asymmetry, and spatio-temporality in action. Adaptive Behavior , 17(5):367--386

    work page 2009

  7. [7]

    Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jackson, T., Jesmonth, S., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, K.-H., Levine, S., Lu, Y., Malla, U., Manjunath, D., Mordatch, I., Nachum, O., Parada, C., Peralta, J., ...

    Pith/arXiv arXiv 2022

  8. [8]

    Brooks, R. (1986). A robust layered control system for a mobile robot. IEEE Journal on Robotics and Automation , 2(1):14--23

    work page 1986

  9. [9]

    Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence , 47(1):139--159

    work page 1991

  10. [10]

    Maruyama, N., Yoshida, T., Sato, H., Masumori, A., Johnsmith, and Ikegami, T. (2025). A concurrent modular agent: Framework for autonomous llm agents

    work page 2025

  11. [11]

    McInnes, L., Healy, J., Saul, N., and Großberger, L. (2018). Umap: Uniform manifold approximation and projection. Journal of Open Source Software , 3(29):861

    work page 2018

  12. [12]

    Minsky, M. (1986). The society of mind . Simon & Schuster, Inc., USA

    work page 1986

  13. [13]

    K., Kim, J., Baghdadi, H., Johnson, B

    Mishra, A. K., Kim, J., Baghdadi, H., Johnson, B. R., Hodge, K. T., and Shepherd, R. F. (2024). Sensorimotor control of robots mediated by electrophysiological measurements of fungal mycelia. Science Robotics , 9(93):eadk8019

    work page 2024

  14. [14]

    Mon-Williams, R., Li, G., Long, R., Du, W., and Lucas, C. G. (2025). Embodied large language models enable robots to complete complex tasks in unpredictable environments. Nature Machine Intelligence , 7(4):592--601

    work page 2025

  15. [15]

    S., O'Brien, J., Cai, C

    Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., and Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , UIST '23, New York, NY, USA. Association for Computing Machinery

    work page 2023

  16. [16]

    Sareen, H., Zheng, J., and Maes, P. (2019). Cyborg botany: Augmented plants as sensors, displays and actuators. In Extended Abstracts of the 2019 CHI Conference on Human Factors in Computing Systems , CHI EA '19, page 1–2, New York, NY, USA. Association for Computing Machinery

    work page 2019

  17. [17]

    Takata, R., Masumori, A., and Ikegami, T. (2024). Spontaneous emergence of agent individuality through social interactions in large language model-based communities. Entropy , 26(12)

    work page 2024

  18. [18]

    Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., and Anandkumar, A. (2023). Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv: Arxiv-2305.16291

    Pith/arXiv arXiv 2023

  19. [19]

    Yoshida, T., Baba, S., Masumori, A., and Ikegami, T. (2024). Minimal self in humanoid robot “alter3” driven by large language model. volume ALIFE 2024: Proceedings of the 2024 Artificial Life Conference of Artificial Life Conference Proceedings , page 53

    work page 2024

  20. [20]

    Yoshida, T., Masumori, A., and Ikegami, T. (2023). From text to motion: Grounding gpt-4 in a humanoid robot "alter3"

    work page 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.