Link to the code: Zae-Project / arkspace-core

How Satellites Negotiate Tasks: Distributed AI Coordination Protocols in Orbit

A constellation of 2,800 satellites cannot be scheduled from the ground in real time. The communication latency alone makes it impossible: a ground station sends an observation request, waits for the round-trip signal delay, receives a response, computes an optimized assignment across thousands of orbital geometries and resource constraints, and sends back commands. By the time the schedule arrives, the observation window has passed.

China’s Three-Body Computing Constellation targets exactly that scale. So does SpaceX’s orbital data center proposal. So does any Earth observation constellation trying to respond dynamically to weather events, maritime traffic, or disaster response needs. The scheduling problem that these constellations face is not a software optimization challenge. It is an architectural one: coordination has to move from the ground to the spacecraft themselves.

The Contract Net Protocol

The Contract Net Protocol (CNP) was described by Reid G. Smith in a 1980 paper in Communications of the ACM. The idea is a market mechanism for task allocation among distributed agents:

  1. A task becomes available (a manager satellite has an observation request it cannot fulfill alone).
  2. The manager broadcasts an announcement describing the task requirements.
  3. Eligible satellites evaluate the request against their current schedule, capabilities, and orbital position, and submit bids.
  4. The manager awards the task to the most qualified bidder.
  5. The winning satellite executes the task and reports results.

This mechanism has no central coordinator. Each satellite acts as both manager (announcing tasks it originates) and contractor (bidding on tasks announced by others). The ground station can inject requests, but does not need to compute the full schedule. The constellation self-organizes.

The reason CNP has persisted for 44 years is that it handles failure gracefully. If a winning satellite fails to execute, the task can be re-announced. If a satellite’s schedule changes due to a thruster anomaly, it withdraws from active bids. The protocol degrades gracefully under the kind of partial failures that are normal in orbital operations.

The 2026 IJAE Paper: Improving CNP for Satellite Constellations

The International Journal of Aerospace Engineering 2026 issue published “An Improvement of Contract Net Protocol for Distributed Satellite Collaborative Task Planning,” addressing specific failure modes that basic CNP does not handle well.

The core problem: task allocation conflicts. In a large constellation, multiple manager satellites may announce overlapping observation tasks simultaneously. Without coordination between them, the same high-capability contractor satellite might win multiple conflicting bids, creating a schedule that is infeasible at execution time.

The 2026 improvement introduces a secondary allocation strategy with a hierarchical disturbance propagation mechanism. When a primary allocation fails or conflicts are detected, the secondary allocation reruns on the remaining unscheduled tasks, reinserting them into the constellation’s available capacity without requiring a full reschedule from scratch. Computational experiments in the paper show significantly higher observation benefit and lower task loss rates compared to unmodified CNP.

The TRL for this approach is 3–4: the algorithmic improvements are analytically validated, and the CNP framework itself has been tested in simulation environments. The gap to flight is in the onboard compute and communication architecture required to run CNP in real time across hundreds of simultaneous satellite nodes.

NASA FAME: 60 Spacecraft, 2026

The most significant near-term validation of distributed multi-agent satellite AI is not a paper. It is the NASA FAME (Formation for Autonomous Mission Execution) mission, beginning in 2026 with over 60 participating spacecraft.

FAME uses two components from NASA JPL’s AI group:

DCOSP (Dynamic Multi-Satellite Constellation Observation Scheduling Problem): a formal problem formulation that models integrated scheduling and execution as a Dynamic Distributed Constraint Optimization Problem (DDCOP). DCOSP captures the full complexity of the scheduling problem: spacecraft dynamics, resource constraints (battery, data storage, communication windows), observation priorities, and the uncertainty of dynamic Earth events.

D-NSS (Dynamic Incremental Neighborhood Stochastic Search): the algorithm that solves DCOSP in real time on-orbit. D-NSS is an incomplete online algorithm, meaning it does not guarantee an optimal solution, but produces good solutions fast enough to respond to changing conditions. When a dynamic event occurs (a new observation request, a satellite failure, or a weather window closing), D-NSS repairs the existing schedule rather than recomputing from scratch.

The 60+ spacecraft scale of FAME makes it the largest in-space demonstration of distributed multi-agent AI in history. Results from FAME will directly inform the architecture of future Earth observation constellations, and the same coordination framework is directly applicable to computing constellations where satellites need to negotiate which node processes a given task.

Deep Reinforcement Learning as a Complementary Approach

CNP and its variants are protocol-based: satellites follow explicit rules for announcement, bidding, and awarding. A parallel research direction uses deep reinforcement learning (DRL) to train coordination policies that emerge from optimization rather than explicit rules.

The Attention-based Distributed Satellite Mission Planning (ADSMP) algorithm uses a neural network with attention mechanisms trained via DRL. Each satellite runs the same policy independently, with attention over the current state of neighboring satellites and pending tasks. The result is a distributed policy where each satellite makes locally optimal decisions that aggregate to near-globally optimal schedules.

The 2026 arXiv preprint “Large-Scale Continual Scheduling and Execution for Dynamic Distributed Satellite Constellation Observation Allocation” (arxiv.org/abs/2601.06188) extends this to continuous operation: the constellation reschedules dynamically as new requests arrive, rather than solving a batch scheduling problem once per orbit.

These DRL approaches require more onboard compute than CNP. Running an attention neural network at inference requires hardware that Carnegie Mellon’s radiation-hardened neuromorphic processors are being specifically designed to provide in radiation-tolerant form. The computational requirements are why the autonomy research profiled in Autonomy Ascending is relevant: the same autonomous guidance and control infrastructure that manages orbital maneuvers also hosts the task coordination layer.

The Connection to Orbital Computing

Distributed task planning for Earth observation constellations and distributed task planning for orbital computing constellations are structurally the same problem.

In an Earth observation constellation, the resource being allocated is imaging time and downlink bandwidth. In an orbital computing constellation, the resource being allocated is compute cycles, memory, and inter-satellite link bandwidth. The CNP announcement-bidding-awarding mechanism applies to both. A ground client submits a compute job. The job is announced as a task. Satellite nodes with available compute capacity bid, providing estimates of execution time and data transfer cost. The winning node executes the job and returns results.

The LLM-based satellite management system at Thales Alenia Space shows what the ground-side management layer looks like when AI is applied to component data and configuration. Distributed task planning is what that system looks like when the AI moves fully on-orbit and manages real-time job allocation.

Edge AI satellite operators like D-Orbit already run inference pipelines on individual satellites. The coordination protocol step is what enables those individual inference-capable nodes to function as a fleet rather than as independent units.

Path Forward

The 2026 IJAE improvement to CNP and the NASA FAME mission represent two different points on the deployment trajectory for distributed satellite AI coordination. CNP improvements are algorithmic contributions that will inform future constellation software stacks. FAME is a real mission that will produce operational data on how multi-agent coordination behaves in the orbital environment under actual hardware constraints.

The gap between current implementations and the coordination required for a 2,800-satellite computing constellation is real. Sixty satellites is not 2,800. But FAME establishes proof of concept at a scale that single-satellite or ground-controlled demonstrations cannot. Each doubling of the satellite count in a distributed coordination framework tests whether the protocol’s complexity scales linearly or exponentially.

If CNP-based coordination scales acceptably to hundreds of satellites, the architectural argument for autonomous constellation operations strengthens considerably. Ground-in-the-loop scheduling, where every observation request requires human approval and ground-computed assignment, cannot match the response time that orbital AI applications require. The protocols being refined in 2026 are what makes autonomous coordination at scale technically plausible.

Official Sources