PLanTS: Plant-Centered Sensing & Control

• PLanTS is a plant-centered research space where living plants and plant-inspired systems serve as active substrates for sensing, control, and fabrication.
• It spans bio-hybrid architecture, artificial morphogenesis, and computational plant representation to foster innovative, responsive systems.
• Research in PLanTS employs decentralized control, advanced sensing modalities, and 3D segmentation to bridge biological insights and engineering applications.

Searching arXiv for the cited papers and topic framing. Searching arXiv for "PLANTS planning-like task summarization (Pallagani et al., 2024)". “PLanTS” (Editor’s term) denotes a plant-centered research space in which living plants, plant-inspired morphogenetic systems, and computational plant representations are treated as active substrates for sensing, control, fabrication, or analysis. In the cited literature, this space includes bio-hybrid architecture coupling natural plants and distributed robots, artificial braided structures governed by decentralized plant-like control, multimodal measurement of plant electrical and kinematic response, mechanistic models of posture control and fluid-structure interaction, biomagnetic structures in vascular tissue, and public datasets for 3D plant segmentation. A distinct 2024 work uses the orthographically similar acronym PLANTS for “planning-like task summarization,” which is methodologically unrelated to botanical systems but relevant to the nomenclature of the term (Hamann et al., 2017, Hofstadler et al., 2018, Gorobets et al., 2019, Mertoğlu et al., 2024, Pallagani et al., 2024).

1. Scope, terminology, and research strands

The cited corpus does not define PLanTS as a single standardized framework. It instead suggests a family of technical programs organized around the proposition that plants are not merely passive biological background, but active material, active sensor, active controller, or active model. This interpretation is supported by works in living architecture, artificial morphogenesis, plant-response instrumentation, theoretical biophysics, and geometric data annotation (Hamann et al., 2017, Hofstadler et al., 2018, Duerr et al., 2020, Oliveri et al., 2023, Mertoğlu et al., 2024).

Inferred strand	Technical focus	Representative source
Bio-hybrid architecture	Living plants with distributed robots, braided scaffolds, and growth steering	(Hamann et al., 2017)
Artificial plant morphogenesis	Braided structures with decentralized Vascular Morphogenesis Controller	(Hofstadler et al., 2018)
Plant observability	Electrophysiology, video tracking, sap flow, transpiration, and plant-response sensing	(Hamann et al., 2017, Duerr et al., 2020)
Plant biophysics	Biomagnetic nanoparticles, morphoelastic posture control, and fluid-structure interaction	(Gorobets et al., 2019, Oliveri et al., 2023, Gosselin, 2019)
Computational plant representation	Annotated 3D point clouds and semantic segmentation	(Mertoğlu et al., 2024)
Acronymically related but non-botanical work	Summarization of planning-like tasks	(Pallagani et al., 2024)

A common misconception is to treat all plant-centered technical work as either biomimetics or horticultural sensing. The corpus is broader. Some papers study living plants directly; some construct physically embodied artificial “plants”; some formalize plant mechanics mathematically; and some build machine-learning benchmarks over plant geometry. The homonymous PLANTS benchmark further shows that the acronym itself is not semantically stable across fields (Pallagani et al., 2024).

2. Living plants as active material in architecture

In flora robotica, living plants are treated as active material for architecture rather than as decoration or post hoc greening. The project proposes a bio-hybrid system in which natural plants and distributed robots are tightly coupled to grow architectural artifacts and spaces over time. The stated motivation is to move beyond orthodox building practice toward systems exhibiting self-repair, material accumulation, and self-organization, with user-defined objectives and occupant interaction steering where growth is desired or prohibited (Hamann et al., 2017).

Growth steering is implemented primarily through light, hormones, and mechanical/environmental stimulation. Blue light is used as an attraction stimulus through phototropism, while far-red light is used as a repelling stimulus through the shade-avoidance syndrome. The paper also identifies auxin-like treatments as a means to alter curvature, growth direction, and architecture, and mentions mechanical stimulation such as vibration as another control modality. These stimuli are embedded in a closed-loop system that also measures plant state through IR-proximity sensing, electrophysiology, and sap flow / transpiration measurements (Hamann et al., 2017).

The sensing architecture is unusually explicit. The MU3.3 phytosensing platform includes two channels for electrophysiological measurements, two separate channels for impedance measurements, two separate channels for bio-potential measurements, a transpiration sensor mounted on the back side of a leaf, temperature, humidity, ambient light, a 3D accelerometer, a 3D magnetometer, and three actuator channels. This suggests a conception of PLanTS systems as cyber-physical assemblies in which plant state is observed directly rather than inferred only from environmental variables (Hamann et al., 2017).

A defining organizational motif is braiding. Braids are produced by robots from continuous material filaments and serve simultaneously as scaffolds for climbing plants, initial architectural artifacts before sufficient plant growth, carriers for robots and electronics, and reconfigurable growth substrates. The modular braiding robot uses driver modules and switch modules arranged into strings, circles, or matrices on a flat surface to move material dispensers through intersecting pathways. The anticipated demonstrator is a small wall with two holes, one functioning as a window where growth remains prohibited and another representing damage that the system must regrow while preserving the window opening (Hamann et al., 2017).

3. Artificial morphogenesis and decentralized vascular control

Where flora robotica couples living plants to robots, “Artificial Plants - Vascular Morphogenesis Controller-guided growth of braided structures” develops a physically embodied but non-living morphogenetic system. Its central problem is ex-silico morphological adaptation through material accumulation: taking plant-inspired growth beyond simulation and into a real braided structure that can continue to “grow” through material addition (Hofstadler et al., 2018).

The key control law is the Vascular Morphogenesis Controller (VMC). Each controller node maintains local state in the form of Resource $R$, Success $S$, and Vessels $V_1,V_2,\ldots,V_n$. The architecture is decentralized: each node updates state locally using sensor data at that node, incoming values from neighboring nodes, and fixed genomic parameters shared by all nodes. The information flow is plant-inspired. Leaf or apical regions compute success from local conditions; $S$ is sent upward; parent nodes aggregate incoming $S$; the result modifies vessel thickness $V$; and $V$ biases the downward allocation of resource $R$. The paper explicitly interprets this loop as a digital analogue of canalization and apical dominance (Hofstadler et al., 2018).

The embodied substrate is a Y-shaped braided module made of PET strapping. Each module has one root node and, in the reported hardware, two leaves that serve as candidate growth sites. Local sensing is performed at the leaves by an accelerometer for posture or tilt and four photoresistors for ambient light. Neighbor communication between Raspberry Pis uses one transmit line, one receive line, and ground. Growth itself is not automatic: the controller identifies the most promising leaf, and a human operator physically braids a new module onto that leaf. The resulting system combines autonomous distributed decision-making with manual realization of growth (Hofstadler et al., 2018).

The controller update logic is summarized as a five-step local loop: receive $R$ from connected parents or generate it; receive $S$ from children; adjust vessel thickness $S$0 according to received $S$1; distribute $S$2 to children according to relative $S$3; and distribute $S$4 to parents according to the amount of $S$5 received. The reported parameter set is $S$6, $S$7, $S$8, $S$9, $V_1,V_2,\ldots,V_n$0, $V_1,V_2,\ldots,V_n$1, $V_1,V_2,\ldots,V_n$2, and $V_1,V_2,\ldots,V_n$3. This formalization places PLanTS work squarely within distributed embodied control rather than mere morphological analogy (Hofstadler et al., 2018).

4. Plant sensing, electrophysiology, and human or acoustic interaction

A distinct PLanTS strand treats plants as measurable responsive systems. In “Eurythmic Dancing with Plants”, plant response is measured through two synchronized channels: electrical activity via a Plant SpikerBox and visible motion via camera-based tracking. The computer-vision pipeline uses Shi-Tomasi Corner Detection and Lucas-Kanade optical flow. For electrical analysis, the authors extract 20 Mel-Frequency Cepstral Coefficients (MFCCs) and report that a 21-second rolling window provides the best tradeoff between sample size and correlation strength (Duerr et al., 2020).

The study reports three experiments. The first compares simultaneous SpikerBox measurements on different plants during the same event and gives average correlations of 0.509 for beetroot, 0.703 for lettuce, and 0.363 for tomatoes. The second compares dancer hand movement to plant electrical discharge and reports weak average correlations, with 0.0412 for the left hand and -0.1051 for the right hand, although some individual coefficients are substantial and statistically marked. The third compares regularly danced plants with first-time exposed plants, reporting averages of 0.471 for the control group and 0.371 for the experimental group. The paper interprets these results as exploratory evidence that plants exhibit measurable signal changes during eurythmic interaction, while also listing limitations including small sample size, possible vibration effects, weather effects, and incomplete microclimate control (Duerr et al., 2020).

A more mechanistic signal-processing account is given in “An Acoustic Communication Model in Plants.” This work develops an end-to-end communication-theoretic pipeline from underground water-flow sound through soil propagation, cell-wall mechanics, MCA2 mechanosensitive channels, a $V_1,V_2,\ldots,V_n$4-ROS hub, CPK29, PIN2, auxin redistribution, and finally asymmetric root bending. Its principal simulation result is that a $V_1,V_2,\ldots,V_n$5 Hz, $V_1,V_2,\ldots,V_n$6 acoustic stimulus elevates cytosolic $V_1,V_2,\ldots,V_n$7 from $V_1,V_2,\ldots,V_n$8 nM to $V_1,V_2,\ldots,V_n$9 nM within 50 s, after which calcium stabilizes in the 220–240 nM range. In the communication-theoretic formulation, the decision rule is based on an Activated PIN2 Ratio with $S$0 if $S$1, and the bit-error-rate analysis indicates that the system needs about 100 s to produce a reliable decision (Merdan et al., 30 Nov 2025).

Taken together, these studies suggest two distinct epistemic styles within PLanTS. One is exploratory and multimodal, using synchronized vision and electrophysiology in a real-world biodynamic setting. The other is mechanistic and model-based, converting hypothesized plant auditory behavior into a layered receiver model. A plausible implication is that plant-response research in this space oscillates between phenomenological instrumentation and communication-theoretic formalization (Duerr et al., 2020, Merdan et al., 30 Nov 2025).

5. Internal biophysics, posture regulation, and flow-driven mechanics

One biophysical line of work studies biogenic magnetic nanoparticles (BMNs) in plant tissue. Comparative genomics is used to argue that all 55 investigated plants with genomes deciphered by more than 50% were potential BMN producers, and AFM/MFM experiments on Nicotiana tabacum, Solanum tuberosum, and Pisum sativum report BMNs in chains located in the wall of phloem sieve tubes within the vascular tissue. Reported BMN sizes are approximately 110–220 nm in tobacco leaf, 80–185 nm in tobacco root, 60–120 nm in potato stem, 35–60 nm in potato tuber, and 95–105 nm in pea stem; chain lengths range from 2 to 10 particles depending on organ. The authors argue that this repeated localization implies common metabolic functions and that BMN chains generate stray gradient magnetic fields of several thousand Oe, with gradients on the order of $S$2–$S$3, sufficient to affect mass transfer near vesicles, granules, liposomes, organelles, membrane structural elements, and amyloplasts (Gorobets et al., 2019).

The same paper reports perturbation experiments with artificial magnetite nanoparticles added to soil at 0.1 mg/ml and 1 mg/ml. In pea, conglomerates form on sieve tube walls and contain both biogenic BMNs and artificial magnetite; at higher concentration, additional and even parallel chains appear. At 0.1 mg/ml, the reported growth responses are a 34% increase in stem length, 22% increase in plant length, and 287% increase in the number of lateral roots. The mechanistic interpretation is that added magnetite modifies the local field organization near BMN chains and thereby alters vesicle or granule transport in phloem-associated regions (Gorobets et al., 2019).

A second line formalizes plant posture control with morphoelastic rods. In “Active shape control by plants in dynamic environments,” the local tropic law is

$S$4

with bending number $S$5. Without rotation, the paper identifies a universal planar simple caulinoid after appropriate rescaling. Under clinostat rotation, it derives a stable family of three-dimensional dynamic equilibria with fixed centerline in space and material rotating around that centerline, and shows linear stability for $S$6 over the tested range $S$7. When axial growth is added, the model predicts steady behaviors described as solitary waves in a co-moving frame (Oliveri et al., 2023).

A third line studies plants as fluid-structure interaction systems. “Mechanics of a Plant in Fluid Flow” organizes the subject through reconfiguration, in which bending, streamlining, and effective velocity reduction lower load relative to a rigid body. The review emphasizes the Reynolds number

$S$8

the Cauchy number

$S$9

the reconfiguration number $S$0, and the empirical Vogel exponent in $S$1. It further extends the framework to poroelasticity, torsion, chirality, buoyancy, skin friction, wave action, flutter, and vortex-induced vibrations, linking mechanics to biomass production, crop lodging, pollen release, mass and heat exchange, and coastal protection (Gosselin, 2019).

6. Geometric datasets, semantic segmentation, and the PLANTS acronym collision

On the representational side, PLANesT-3D is a public annotated dataset of 34 RGB 3D point clouds representing 34 real plants from three species: 10 Capsicum annuum, 10 Rosa kordana, and 14 Ribes rubrum. The data are reconstructed from Structure from Motion (SfM) and Multi-View Stereo (MVS) using images captured by a handheld DSLR camera at 6240 × 4160 resolution. Point clouds are preprocessed through scale recovery, pose normalization with MSAC, and plant extraction by removing negative-$S$2 points and retaining the largest connected component. The annotation scheme provides semantic labels for leaf and stem, plus organ instance labels in which each leaflet receives its own ID while all stem points share a single instance label (Mertoğlu et al., 2024).

The paper also introduces SP-LSCnet, a two-stage semantic segmentation method combining unsupervised superpoint extraction and a PointNet++-style classifier enhanced with the Center Shift Module (CSM) and Radius Update Module (RUM). The dataset is split species-wise into 70% training and 30% testing. Segmentation is evaluated with Precision, Recall, per-class IoU, Accuracy, and MIoU. Reported MIoU values are 94.3 for PointNet++, 95.5 for RoseSegNet, and 95.0 for SP-LSCnet on pepper; 89.8, 92.0, and 89.8 on rose; and 94.2, 95.1, and 94.5 on ribes. The paper concludes that RoseSegNet achieved the highest overall accuracy and mean IoU, while SP-LSCnet was competitive with PointNet++ and produced smoother segmentations because it classifies superpoints rather than individual points (Mertoğlu et al., 2024).

A terminological complication arises because PLANTS also names a non-botanical summarization benchmark. “PLANTS: A Novel Problem and Dataset for Summarization of Planning-Like (PL) Tasks” defines plan summarization over a set of action sequences for a common goal,

$S$3

and constructs a dataset spanning automated plans, recipes, and travel routes. The dataset contains 130 plans, with 10 different problems/goals per domain; automated plans are generated with SymK using $S$4, recipes are selected from Recipe1M+, and routes are obtained from the OpenStreetMap API. Baselines include GPT-4o, TextRank, and a frequency-based extractive method. In a user study with 10 annotators, Cohen’s kappa is 0.72, and GPT-4o is rated easiest to understand across all three tasks, but the authors explicitly warn that readability does not guarantee preservation of executional semantics and identify hallucination as a concern. This work is relevant chiefly because it fixes the orthographic distinction: not every “PLANTS” paper concerns plants in the botanical sense (Pallagani et al., 2024).

The coexistence of PLANesT-3D and PLANTS underscores that PLanTS is best understood contextually. In botanical and bio-hybrid research, the term points toward plant material, plant-inspired control, or plant observability. In NLP, the same letter string may denote an unrelated benchmark over action sequences. For technical reading, disambiguation by domain and citation is therefore essential (Mertoğlu et al., 2024, Pallagani et al., 2024).