iSatCR: Graph-Empowered Joint Onboard Computing and Routing for LEO Data Delivery

This paper presents iSatCR, a distributed graph-based deep reinforcement learning framework that jointly optimizes onboard computing and routing in Low Earth Orbit satellite networks to alleviate bandwidth bottlenecks by processing data in situ and significantly reducing transmission volume.

The Gist

Imagine a massive fleet of thousands of satellites circling the Earth, acting like a giant, floating camera crew. Every day, they snap millions of high-definition photos and videos of our planet—tracking storms, monitoring crops, and watching cities grow.

The Problem: The "Traffic Jam" in the Sky
Here's the catch: These satellites are generating so much data (about 1 Terabyte per satellite per day!) that they can't just beam it all down to Earth fast enough. The connection to the ground is like a narrow straw trying to drink from a firehose. If they try to send everything raw, the data gets stuck, delayed, or lost.

Traditionally, engineers tried to fix this by building better "roads" (routing) to move the data faster. But the traffic was just too heavy.

The Solution: "Cooking" the Data in Space
The authors of this paper, iSatCR, propose a brilliant new idea: Don't send the raw ingredients; send the cooked meal.

Instead of sending the massive raw video files down to Earth, the satellites should process the data while they are in space.

• Raw Data: A 50GB video file of a forest fire.
• Processed Data: A 5KB text message saying, "Fire detected at these coordinates."

By doing the "cooking" (computing) in space, the amount of data sent to Earth shrinks dramatically. But this creates a new, tricky puzzle: Who cooks what, and how do they pass the ingredients?

The Challenge: The Cosmic Kitchen
Imagine a kitchen with 1,000 chefs (satellites) moving at 17,000 mph.

1. They are always moving: The "kitchen" layout changes every second as they orbit.
2. Resources are uneven: Some chefs have powerful ovens (computers) but no counter space (storage). Others have plenty of space but weak ovens.
3. Communication is hard: They can't all talk to the "Head Chef" on Earth instantly because the signal takes too long. They have to decide on the fly.

If a satellite tries to cook a task but runs out of space, it has to pass the raw data to a neighbor. But which neighbor? The one with the best oven? The one closest to the ground? The one with the shortest line?

The iSatCR Solution: The "Smart Neighborhood" Network
The paper introduces iSatCR, a system that lets the satellites act like a super-smart neighborhood.

1. The "Shifted Feature" Gossip (Graph Embedding)

In a normal network, a satellite might only ask its immediate neighbor, "Are you busy?"
iSatCR uses a technique called Graph Embedding. Think of this as a satellite asking its neighbor, "Who is your neighbor, and what are they doing?"

• The Analogy: Imagine you are at a party. Instead of just asking the person next to you if the bathroom is free, you ask them, "Who is standing near the bathroom, and are they blocking it?"
• The "Shift": The system organizes this information into layers (1-hop, 2-hop, 3-hop neighbors). It creates a mental map of the "kitchen" that extends three steps away. This allows a satellite to see not just who is right next to it, but who is down the line, helping it avoid traffic jams before they happen.

2. The "Smart Brain" (D3QN)

Once the satellite has this map, it needs to make a decision. Should I cook this myself? Should I pass it to the left? To the right?
The system uses Deep Reinforcement Learning (DRL), specifically a "Dueling Double Q-Network."

• The Analogy: Think of this as a video game AI that has played the game millions of times. It learns by trial and error.
  • If it passes the data to a satellite that is overloaded, it gets a "penalty" (slow speed).
  • If it finds a satellite with a fast oven and a clear path to the ground, it gets a "reward."
• Over time, the satellites learn the perfect strategy to balance the load, ensuring no single satellite gets overwhelmed while others sit idle.

3. The "Heuristic" Shortcut

To make the AI learn faster, the authors added a "cheat code" (Heuristic Exploration).

• The Analogy: When the AI is learning, it usually tries random moves. But sometimes, it's told: "If you've already cooked the food, just pick the path that looks shortest to the exit." This helps the system learn the right moves much quicker.

The Results: Why It Matters
The authors tested this in a massive simulation with thousands of satellites.

• Speed: iSatCR was much faster than the old methods, especially when the network was crowded (high load).
• Reliability: Even when satellite links failed (like a road closing due to a storm), iSatCR found new routes quickly without losing data.
• Fairness: It balanced the work perfectly, so no single satellite was doing all the heavy lifting while others did nothing.

In Summary
iSatCR is like giving every satellite in a massive fleet a shared, real-time map of the entire neighborhood and a super-intelligent brain that knows exactly how to pass tasks around to avoid traffic jams. Instead of clogging the pipes to Earth with raw data, it processes the information in space and sends down only the valuable results, making our Earth observation system faster, smarter, and more efficient.

Technical Summary

1. Problem Statement

Low Earth Orbit (LEO) satellite constellations are increasingly used for high-resolution Earth observation, generating massive data volumes (up to 1 TB per satellite per day). Transmitting this raw data to ground stations creates a severe bottleneck due to limited bandwidth and uneven ground station distribution.

• Limitations of Current Approaches: Traditional routing-only strategies cannot handle the surging data volume. While onboard computing (processing data in situ to reduce volume) offers a solution, jointly optimizing routing (transmission paths) and computing (where to process data) is challenging due to:
  1. High Overhead: Frequent collection of global network states (topology and resource loads) in a highly dynamic environment (high-speed movement, link failures) incurs significant communication overhead.
  2. Dynamic Resource Distribution: Computing resources are unevenly distributed; data often needs offloading to other satellites, requiring real-time awareness of neighbor capabilities.
  3. Complexity: The joint optimization problem involves high-dimensional state and action spaces, leading to the "curse of dimensionality" in large-scale constellations.

2. Methodology: iSatCR

The authors propose iSatCR, a fully distributed framework that jointly optimizes onboard computing and routing using Graph-empowered Deep Reinforcement Learning (DRL). The system is modeled as a Partially Observable Markov Decision Process (POMDP).

A. Distributed Resource Awareness (Graph Embedding)

To address the overhead of global state collection, iSatCR uses a novel graph embedding mechanism:

• Shifted Feature Aggregation: Instead of standard aggregation, the method aggregates features from 1-hop and 2-hop neighbors and "shifts" them into distinct regions of the embedding vector. This preserves spatial information regarding resource distribution (e.g., distinguishing between immediate neighbors and those further away).
• Distributed Message Passing: Satellites exchange graph representations with neighbors. The embedding captures the current satellite's state, aggregated 1-hop features, and aggregated 2-hop features, effectively providing a 3-hop perception range.
• Fault Handling: Special padding techniques are used to mark unavailable neighbors or link failures, ensuring the model remains robust to topology changes.

B. Joint Decision Making (D3QN)

The decision-making core utilizes a Dueling Double Deep Q-Network (D3QN):

• State Space: Includes local resource states (storage, queue lengths), graph embeddings of neighbors (3-hop range), task requirements (data size, compute demand), and destination direction.
• Action Space: Decides whether to offload the task to a specific neighbor's transmission queue or process it locally in the computing queue.
• Reward Function: Designed to minimize total task delay (transmission + propagation + queuing + computing) while penalizing packet loss and storage threshold violations.
• Heuristic Exploration: A hybrid exploration strategy is employed. During training, if a task is already computed, the agent prioritizes transmission actions; if not, it considers both. A heuristic component guides exploration based on shortest-hop paths to accelerate convergence.

3. Key Contributions

1. Novel Graph Embedding Mechanism: Introduced a "shifted feature aggregation" method that allows satellites to efficiently sense resource states within a 3-hop range without global flooding, significantly reducing communication overhead while preserving spatial resource distribution details.
2. Distributed D3QN Framework: Proposed a graph-empowered D3QN algorithm tailored for the POMDP nature of LEO networks. It handles the high-dimensional state/action space and adapts to dynamic topologies better than centralized approaches.
3. System-Level Simulation: Developed a comprehensive simulation environment modeling the constellation scale (up to thousands of satellites), including realistic link failure models (Markov process), task generation, and storage constraints.
4. Performance Validation: Demonstrated that the proposed method outperforms baselines (DDQN, PPO, D3QN without graphs, and an Ideal Centralized Solution) in terms of delay, packet loss, and load balancing, particularly under high-load and high-failure scenarios.

4. Experimental Results

The authors evaluated iSatCR against four baselines across different constellation sizes (Iridium, Telesat, OneWeb, Starlink) and varying conditions:

• High-Load Performance: Under high task loads (>85 tasks/sec), iSatCR maintained significantly lower average delays and packet loss rates compared to centralized and other DRL methods. The graph embedding allowed it to find optimal computing nodes closer to the source, reducing transmission hops.
• Robustness to Link Failures: When link failure rates increased, iSatCR showed the most stable performance with the lowest packet loss and minimal increase in transmission hops, proving its adaptability to dynamic topology changes.
• Load Balancing: The cumulative distribution of computing time showed iSatCR achieved the most balanced load distribution across satellites, preventing hotspots.
• Generalization: The algorithm maintained superior performance across different constellation scales (from Iridium to Starlink), indicating strong generalization capabilities without retraining.

5. Significance

This paper addresses a critical bottleneck in next-generation LEO satellite networks: the efficient delivery of massive Earth observation data. By shifting from "routing-only" to "joint computing and routing," iSatCR leverages onboard processing to drastically reduce downlink bandwidth requirements.

• Scalability: The distributed nature of the solution makes it scalable to mega-constellations where centralized control is computationally infeasible.
• Efficiency: The graph-embedding approach provides a middle ground between local-only decisions (which lack context) and global flooding (which is too heavy), enabling intelligent, context-aware decisions with low overhead.
• Practicality: The results suggest that iSatCR can significantly improve the timeliness and reliability of Earth observation data delivery, which is vital for applications like disaster recovery and environmental monitoring.