Jasmine: Cross-Domain Technical Systems

Updated 3 July 2026

Jasmine is a suite of diverse systems spanning microrobot swarms, space missions, stellar dynamics, Arabic language models, data visualization, and cybersecurity frameworks.
It enables scalable experiments and precise measurements through innovations in hardware, algorithmic strategies, and distributed computing across multiple research domains.
The systems integrate advanced sensing, communication, and computational techniques, achieving high reproducibility, cost effectiveness, and robust performance in practical applications.

Jasmine refers to a diverse set of technical systems in robotics, astrophysics, language modeling, cyber-physical infrastructure, and scientific computing. Described below are key Jasmine instances from research literature, spanning microrobotic swarms, infrared astrometric satellites, axisymmetric stellar dynamics modeling, modern deep generative LLMs, cyber-physical data exploration, world modeling codebases, and cybercrime active learning frameworks.

1. Jasmine Microrobot Platform for Swarm Robotics

The Jasmine microrobot is an open-hardware modular platform developed for scalable and reproducible experiments in collective behavior, swarm intelligence, and distributed perception. The hardware consists of a compact (26–30 mm cubic, ~10–15 g) differential-drive chassis with two DC motors, two Atmel AVR microcontrollers (a main ATmega168 at 16 MHz and a motor/IO ATmega88 at 8 MHz), IR-based directional communication, and onboard IR sensing systems for proximity, geometry, and ambient-light detection. The architecture, available via swarmrobot.org, supports swarms of 10–100+ robots operating autonomously with neighborhood coordination, robust shape and object recognition, and real-time information diffusion (Kernbach, 2011, Kornienko, 2011, Kernbach, 2011, Kernbach, 2011).

Communication operates through six or more optically isolated IR emitter-receiver pairs covering 60–90° sectors, enabling localized, half-duplex, on/off-keyed links at ≈1 kb/s with 150 mm range and spatial channel context. Perception combines real-time IR scanning (with 1 mm resolution at 20–70 mm) and distance decoding via empirical ADC calibration. Minimalist behavioral control cycles interleave sensing, IR transmission, navigation, and decision-making at 10–150 ms intervals (Kornienko, 2011). Swarm behaviors include collective perception, region segmentation, bio-inspired aggregation, and multi-hop message diffusion.

Research has demonstrated multiple classes of swarm information transfer: local (stigmergic) interactions, epidemic-style global broadcasts, and explicit feedback coordination. Connectivity and information propagation are characterized by network-theoretic metrics (<k>, D, T_p) empirically measured in arenas of 10–100 robots (Kernbach, 2011). The system is engineered for high experimental reproducibility, low manufacturing cost, and rapid software-hardware iteration.

2. JASMINE: Near-Infrared Space Astrometry Mission

JASMINE (Japan Astrometry Satellite Mission for INfrared Exploration) is a Japanese space mission targeting Gaia-level astrometric precision (σπ ≈ 25 μas, σμ ≈ 25 μas yr⁻¹ at H_w<12.5 mag) in the near-infrared. Its primary aim is 3D mapping of the heavily extincted Galactic Center via repeated short-exposure (12.5 s) imaging, using a 36 cm NIR telescope and four 2k×2k InGaAs arrays (0.472″/pixel, 0.55° FoV), yielding ~60,000 visits per star in three years (Kawata et al., 2023, Kamizuka et al., 2024, Makarov et al., 2012).

Astrometric centroiding is achieved with ePSF fitting at ~4 mas per exposure, co-added to reach the mission targets. Accurate error modeling leverages scalar and vector spherical harmonics for decomposing correlated (zonal) errors, enabling isolation and mitigation of low-order spatial systematics at the <20–25 μas level (Makarov et al., 2012). The time-series NIR photometry (1 mmag for bright sources) enables auxiliary science such as exoplanet transit searches, asteroseismology, variable star mapping, and microlensing. Survey implementation includes both dedicated Galactic Center mapping and exoplanet campaigns on mid-late M dwarfs (sensitive to 0.1%–0.3% transit depths) (Kawata et al., 2023).

Simulations show that attitude control error (ACE) and detector rolling shutter introduce spatially variable, non-Gaussian PSFs. The JASMINE-imagesim framework provides a realistic pipeline for quantifying these systematics and informing both hardware and analysis countermeasures (Kamizuka et al., 2024).

3. JASMINE: Axisymmetric Multi-Component Stellar Dynamics Code

The JASMINE codebase for axisymmetric stellar systems provides fast, flexible Jeans modeling for galaxies composed of multiple structural/dynamical stellar populations, NFW dark halos, and a central black hole (Caravita et al., 2019). For each stellar component, arbitrary axisymmetric density profiles (e.g., Jaffe, Sersic) and multi-population mass-to-light ratios, ages, metallicities, and kinematics (via Satoh decomposition, parameter k) are supported.

The code numerically solves the axisymmetric Jeans equations: $\frac{\partial}{\partial z}\Bigl(\rho_{*i}\,\sigma_i^2\Bigr) = -\,\rho_{*i}\,\frac{\partial\Phi_{\rm tot}}{\partial z}, \quad \frac{\partial}{\partial R}\Bigl(\rho_{*i}\,\sigma_i^2\Bigr) = \frac{\rho_{*i}}{R}\Bigl(\overline{v_\varphi^2}_i-\sigma_i^2\Bigr) - \rho_{*i}\frac{\partial\Phi_{\rm tot}}{\partial R}$ for each population $i$ , where $\Phi_{\rm tot}$ includes all mass components. Light-weighted kinematic observables and stellar-population gradients are readily generated. A “basis” solution strategy allows rapid post-processing of different mass and kinematic parameterizations (Caravita et al., 2019).

4. JASMINE: Arabic GPT Transformer LLMs

The JASMINE suite comprises four autoregressive decoder-only Transformer models, optimized for Arabic (Classical, MSA, dialectal) using a corpus of 235 GB (~23B tokens) and 64k BPE vocabulary (Nagoudi et al., 2022). Model scales span from 350M to 6.7B parameters (12-32 layers). The benchmark covers 23 Arabic datasets for language modeling, autocompletion, commonsense inference (AraSWAG), word manipulation, and NLU.

Zero- and few-shot evaluation shows JASMINE models achieve state-of-the-art perplexity and F₁ relative to AraGPT2 and mGPT, with JASMINE 6.7B reaching 42.25 avg. perplexity and best performance in AraSWAG and ORCA NLU. Human evaluation demonstrates that model generations are fluently indistinguishable from human text at chance rates. Bias audits reveal prevalence of gender, region, and religion-based stereotypes in completions. Mitigation strategies focus on bias auditing, access controls, and monitored deployment (Nagoudi et al., 2022).

5. Jasmine in Scientific Infrastructure, Data Exploration, and World Modeling

a. Jasmine: JAvaScript Multimodal INformation Explorer

Jasmine is also a web-based system for astronomical data exploration, supporting multimodal inspection of image and point cloud data (e.g., Illustris TNG galaxies) (Schweder et al., 30 Apr 2025). It employs a fully client-side architecture using Aladin Lite for images and three.js for interactive point clouds. HiPS-based tiling, autoencoder-based latent embedding (3D VAE), and on-demand detail modals enable scalable, low-overhead scientific visualization. Performance is interactive for ~10⁵ particles and O(logN) complexity for zoom. Extensions for automated window layout, bidirectional coordination, and plugin data pipelines are planned.

b. Jasmine: JAX World Modeling Codebase

The Jasmine JAX-based codebase implements a performant, scalable, and reproducible infrastructure for world modeling across single and multi-host accelerator configurations (Mahajan et al., 30 Oct 2025). The pipeline comprises a high-throughput Grain+ArrayRecord data loader, VQ-VAE video tokenizer, ST-Transformer-based dynamics model, and distributed mixed-precision JAX training with pjit/sharding via Shardy. Benchmarking demonstrates an order-of-magnitude speedup over prior open-source implementations (e.g., Genie in CoinRun), with 6,709 frames/s per GPU throughput and bit-wise reproducibility.

Key features include flexible dataset ingestion, checkpointing (Orbax), code extensibility for new model backbones or sampling, and robust scaling to O(100) accelerators (≥90% efficiency). The system supports a wide range of research in RL and generative modeling.

6. Jasmine in Cybersecurity and Machine Learning

Jasmine, in the context of cybercrime and intrusion detection, is a hybrid active learning (AL) method based on α-dynamic updating of query strategies. Unlike static or fixed-hybrid approaches, Jasmine maintains and updates the weights (αₐ, α_z, αᵣ) for anomaly, uncertainty, and random queries, using error-based feedback to optimize the query mix during labeling rounds (Klein et al., 2021). Mathematical definitions involve margin (uncertainty) and class-conditional anomaly scores.

Empirical evaluation on NSL-KDD and UNSW-NB15 benchmarks shows Jasmine consistently achieves equal or higher F₁ with 3–10× fewer labels compared to baselines (ALADIN, pure anomaly, or uncertainty querying). The scheme robustly adapts to data/label drift and maximizes early learning curve area, demonstrating particular utility for live NIDS settings (Klein et al., 2021).

Table: Jasmine Systems – Key Technical Domains and Distinctions

Jasmine System	Domain	Core Functionality / Scientific Aim
Microrobot Swarm Platform	Swarm robotics	Distributed sensing, IR communication, collective behavior (Kernbach, 2011, Kornienko, 2011, Kernbach, 2011)
JASMINE Space Astrometry Mission	Astrophysics, space	NIR microarcsecond astrometry, transit photometry, complexity-aware error analytics (Kawata et al., 2023, Kamizuka et al., 2024, Makarov et al., 2012)
JASMINE Jeans Modeling Code	Stellar dynamics	Axisymmetric multi-population dynamical modeling under self/gravitating potentials (Caravita et al., 2019)
JASMINE Arabic GPT Models	Natural Language Processing	Large-scale Arabic generative Transformer, few-shot capabilities, bias evaluation (Nagoudi et al., 2022)
Jasmine Web Scientific Explorer	Astronomical visualization	Multimodal (image+cloud), autoencoding, browser-based exploration (Schweder et al., 30 Apr 2025)
Jasmine JAX World Modeling	AI/ML, simulation	Scalable world modeling, JAX-based, distributed, reproducible (Mahajan et al., 30 Oct 2025)
Jasmine Cybersecurity Active Learning	Cyber-ML, security	Dynamic-credit hybrid AL, drift-robust intrusion detection (Klein et al., 2021)

These Jasmine instances exemplify best practices in open-system design, cross-disciplinary engineering, and robustness against hardware, software, or data-related failure modes. The unifying factor is an emphasis on integrating computational, sensing, and communication primitives to enable scalable research across robotics, astrophysics, linguistics, and AI.