Post-Doctoral Research Visit F - M Agentic Network-Server Function Placement For Direct-To-Satellite Iot H/F - INRIA

A propos d'Inria

Inria est l'institut national de recherche dédié aux sciences et technologies du numérique. Il emploie 2600 personnes. Ses 215 équipes-projets agiles, en général communes avec des partenaires académiques, impliquent plus de 3900 scientifiques pour relever les défis du numérique, souvent à l'interface d'autres disciplines. L'institut fait appel à de nombreux talents dans plus d'une quarantaine de métiers différents. 900 personnels d'appui à la recherche et à l'innovation contribuent à faire émerger et grandir des projets scientifiques ou entrepreneuriaux qui impactent le monde. Inria travaille avec de nombreuses entreprises et a accompagné la création de plus de 200 start-up. L'institut s'eorce ainsi de répondre aux enjeux de la transformation numérique de la science, de la société et de l'économie.
Post-Doctoral Research Visit F/M Agentic Network-Server Function Placement for Direct-to-Satellite IoT
Le descriptif de l'offre ci-dessous est en Anglais
Type de contrat : CDD

Niveau de diplôme exigé : Thèse ou équivalent

Fonction : Post-Doctorant

Contexte et atouts du poste

Context

Direct-to-satellite Internet of Things (DtS-IoT) extends low-power terrestrial IoT technologies such as LoRaWAN to remote areas where terrestrial infrastructure is unavailable. In this architecture, resource-constrained sensors transmit directly to Low Earth Orbit (LEO) satellites, which act as spaceborne LoRaWAN gateways. This makes global IoT coverage possible using low-cost devices, but also introduces strong timing, mobility, and visibility constraints [1].

Recent work on the has shown that downlink reliability in satellite LoRaWAN is dominated by timing feasibility: after each uplink, a device opens narrow receive windows, typically one and two seconds after transmission, and any acknowledgment, join response, security message, or application command must arrive within one of them. We address this challenge with a Mixed-Integer Linear Programming (MILP) scheduler named DORSAL-MILP [2] that jointly selects the gateway-to-satellite path and the packet injection time. Its results show that inter-satellite links (ISLs) can dramatically improve ACK delivery by decoupling the feeder satellite; however, DORSAL-MILP still assumes that the LoRaWAN Network Server (NS) logic remains centralized on the ground, creating a fundamental architectural limitation in which decisions about acknowledgments, join procedures, security handling, duplicate suppression, and gateway selection depend on intermittent space-ground connectivity.

This motivates a new architecture in which selected NS functions are distributed to satellites as lightweight, policy-constrained proxy functions. Instead of merely forwarding packets, satellites would reason locally about whether they can answer immediately, forward a downlink scheduling action to a neighbor, defer the decision to the ground NS, or request policy updates when state is ambiguous. This turns direct-to-satellite LoRaWAN into an agentic cyber-physical network: a time-varying constellation of constrained agents that must reason under partial observability, strict protocol rules, orbital dynamics, and security constraints.

This postdoctoral offer is proposed within the framework of the Inria International Relations Department (DRI) 2026 Postdoctoral Campaign with Inria Chile in the domain of Agentic and reasoning AI systems and their applications.

Research Problem

The central research problem and objective of this postdoctoral proposal is:

Which LoRaWAN Network Server functions can be safely and effectively distributed to LEO satellites, and how can agentic AI methods orchestrate their placement and execution under dynamic orbital, communication, energy, and security constraints?

This problem has two coupled dimensions.

1) First, there is an architectural question. LoRaWAN NS functions differ in their tolerance to delay, consistency requirements, statefulness, and dependence on global network knowledge. Some functions may be safe to decentralize, such as immediate preparation of confirmed uplink ACKs, local downlink timing decisions, or temporary buffering. Others may require strict centralization, such as long-term device-session management, encryption key distribution and handling, billing/accounting, or application-level policy enforcement. The project must therefore develop a principled functional decomposition of the NS for DtS-IoT.

2) Second, there is a decision-making question. Even after deciding which functions can be distributed, the system must determine when and where to place them. A satellite may act as a proxy for a given device class, traffic class, region, or time window. These placements should adapt to predicted contact opportunities, ISL connectivity, ground station access, device traffic patterns, onboard compute limits, and the risk of stale state. This requires reasoning and planning rather than simple reactive rules.

The proposed postdoc will approach both dimensions by investigating agentic and reasoning AI techniques for this setting. The goal is not to deploy an unconstrained black-box agent inside the protocol stack, but to design verifiable, policy-aware decision agents that combine optimization, learning, and protocol constraints.

This topic is particularly well aligned with the DRI 2026 Inria Chile call because it combines AI-based reasoning and autonomous agents with a concrete, high-impact application: resilient satellite IoT. It also creates a strong bridge between Inria expertise in DtS-IoT optimization and Chilean expertise in wireless networks, IoT, 5G/6G dynamics, and reinforcement learning.

Mission confiée

Objectives

The project will pursue the following specific objectives:

- Define a functional decomposition of the LoRaWAN Network Server for DtS-IoT, identifying which functions can be centralized, replicated, delegated, cached, or executed by satellite proxies.
- Formalize the function-placement problem over a time-varying LEO constellation with ISLs, ground stations, LoRaWAN Class A timing constraints, onboard compute/storage limits, and security/state-consistency requirements.
- Design an agentic orchestration framework in which satellites act as constrained NS proxy agents that decide whether to acknowledge, forward, defer, buffer, or escalate downlink-related actions.
- Extend the downlink scheduling model to joint function placement, using the DORSAL-MILP model as a theoretical upper-bound benchmark and a source of supervision for faster online policies.
- Develop learning-augmented heuristics or reasoning policies for online operation when full-window observability is unavailable, and MILP computation is too expensive.
- Validate the resulting models and architecture through realistic DtS-IoT scenarios, comparing centralized NS, static proxy placement, optimization-based placement, and agentic adaptive placement. FLoRaSat [6] can be used to this end.

Produce open research artifacts: models, simulation scenarios, datasets, and at least one high-quality publication on agentic distributed NS intelligence for satellite IoT.

Principales activités

Methodology and Work Plan

The work is envisioned as a 12- to 24-month postdoctoral project.

Phase 1: Architecture and Functional Decomposition

The first phase will map the LoRaWAN NS function set to the satellite-IoT context. Candidate decomposed functions include ACK generation, downlink timing, gateway/satellite selection, duplicate-frame handling, ADR policy, join-response preparation, session-state caching, security-context management, buffering, and conflict resolution. Functions will be classified according to statefulness, required state freshness, latency sensitivity, consistency requirements with the central NS, and compute, memory, and communication footprints.

The outcome will be a design space and taxonomy for DtS-IoT NS decomposition, distinguishing functions that are safe to execute locally via a proxy from those that must remain centralized.

Phase 2: Optimization Model for Joint Scheduling and Function Placement

The second phase will extend DORSAL-MILP from a downlink scheduler into a joint decision model. The current DORSAL-MILP selects feasible downlink routes and injection times. The extended model will introduce placement variables representing which satellite hosts which proxy function, for which device/session/region/time horizon, and under which consistency policy. The model will consider time-varying satellite-device, satellite-ground, and satellite-satellite contacts; ISL routing and forwarding delays; ground station contact opportunities; policy synchronization costs between satellite proxies and the central NS; and other factors.

This optimization model will provide two things: an architectural upper bound and a generator of expert decisions for training or evaluating online policies.

Phase 3: Agentic Orchestration and Reasoning Policies

The third phase will design online decision agents for satellites. Each agent will observe a local state: current and predicted contacts, queued uplink/downlink events, local proxy state, known device/session metadata, energy and compute budgets, and recent policy updates from the ground NS.

The agent action space may include: sending a local ACK in RX1/RX2, forwarding an ACK action to a neighboring satellite, requesting or awaiting ground NS confirmation, delegating a proxy role to another satellite, and triggering conflict resolution with the central NS. The project will compare several decision structure approaches: rule-based and protocol-constrained baselines, reinforcement learning for adaptive placement under uncertainty, and learning-to-optimize policies trained from DORSAL-MILP-style optimal solutions. The emphasis will be on safe autonomy: the satellite proxy should not merely maximize delivery, but preserve protocol correctness and degrade gracefully when the local state is insufficient.

Phase 4: Evaluation and Validation

The proposed solution will be evaluated using DORSAL-MILP and, when appropriate, FLoRaSat [6] for end-to-end simulation. Evaluation scenarios will include bent-pipe and ISL-capable constellations, sparse and dense ground station deployments, varying traffic loads, different onboard processing budgets, and imperfect knowledge of orbital/contact predictions.

Key performance metrics will include: downlink ACK success ratio, RX1/RX2 delivery distribution, end-to-end downlink latency, ground-station dependency and control-plane overhead, proxy-state synchronization overhead, robustness to ground-station outages or delayed policy updates, and compute/memory/energy cost per satellite. The comparison will include at least four architectural baselines: a) fully centralized ground NS, b) static satellite proxy placement, c) optimization-based offline placement, and d) agentic adaptive proxy placement.

Expected Outcomes

The expected scientific outcome is a new architecture for distributed LoRaWAN NS intelligence in direct-to-satellite IoT, supported by formal models and simulation evidence. The project should clarify which NS functions can be safely moved into the constellation and quantify the benefit of doing so.

The expected technical outcomes are:

- A taxonomy of LoRaWAN NS functions and their suitability for satellite proxy execution.
- A joint scheduling and function-placement model extending DORSAL-MILP.
- A set of agentic decision policies for satellite NS proxies.
- Open simulation scenarios and reproducible benchmarks.
- A validated comparison between agentic architectures.

One or more publications targeting venues in AI for satellite networks.

Compétences

Technical skills and level required:
Strong programming and research skills in at least one of the following areas: satellite networking, distributed systems, IoT architectures, optimization, artificial intelligence, machine learning, or reinforcement learning. Experience with network simulation frameworks, optimization tools (e.g., MILP/LP solvers), protocol design, or AI-driven decision systems is highly desirable. Familiarity with Python and scientific/research-oriented software development is expected. Knowledge of LoRaWAN, DTN, LEO satellite systems, or wireless networking is appreciated.

Languages:
Excellent written and spoken English required. Knowledge of French or Spanish is appreciated but not mandatory.

Relational skills:
Strong communication and collaboration abilities in international and interdisciplinary environments. Capacity to work autonomously while maintaining active scientific exchanges with supervisors and collaborators. Curiosity, initiative, adaptability, and critical thinking are essential.

Other valued/appreciated:
Experience publishing scientific research, open-source software development, or participation in collaborative research projects. Interest in bridging AI reasoning methods with real-world communication systems and cyber-physical infrastructures is particularly valued.

Avantages

- Subsidized meals
- Partial reimbursement of public transport costs
- Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
- Possibility of teleworking (after 6 months of employment) and flexible organization of working hours
- Professional equipment available (videoconferencing, loan of computer equipment, etc.)
- Social, cultural and sports events and activities
- Access to vocational training
- Social security coverage

Rémunération

€2,788 gross per month