Digest · Services in orbit · read after Chapter 9
導讀 · 在軌服務 · 第 9 章後閱讀
OrbitalBrain: Federated Learning over Inter-Satellite LinksOrbitalBrain：經由星間鏈路的聯邦學習
This digest explains how ISLs change satellite federated learning from waiting for ground passes to routing over a changing graph. OrbitalBrain aggregates updates along a shortest-path tree. Its ~100 Mbps ISL cap sits five orders of magnitude below Part III’s formation-cluster links.讀完後，你能說明為什麼加入星間鏈路會把地球觀測機隊上的聯邦學習，從「等待地面過頂」的問題變成「在一張時隱時現的圖上路由」的問題。OrbitalBrain 如何沿最短路徑樹彙總更新。以及為什麼它約 100 Mbps 的 ISL 上限使它比第三部分的編隊叢集低五個數量級。
Main result
主要結果
OrbitalBrain aggregates satellite learning updates over opportunistic ISLs. Its shortest-path tree increases aggregation opportunities under a small-board power budget and a stated link-capacity model.
OrbitalBrain 經機會式 ISL 彙總衛星學習更新，在小型板卡電力預算及明列鏈路容量模型下，以最短路徑樹增加彙總機會。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| ISL-based aggregation | 2026 · 4–6 · 7 | OrbitalBrain is a NINeS 2026 workshop paper whose group includes three FedSpace authors. It replaces ground-only aggregation with ISL-based aggregation among Earth-observation satellites. Daily ISL opportunities number hundreds to thousands. Ground access offers 4–6 passes, each at most 7 minutes [94, §1]. |
| Capacity and rate | 4 GB · 7.5 W · 7 W · 100 · 10 s · 0.9 | Each satellite has a Jetson Orin Nano 4 GB drawing about 7.5 W, with about 7 W solar input. Median ISL and ground bandwidth is about 100 Mbps, and simulated ISLs are capped at 100 Mbps. Setup above 10 s receives a flat 0.9 efficiency factor [94, §2.1, §4, §5.1, App. B]. |
| Cost assumptions | Qualitative scope | A shortest-path tree roots aggregation at the satellite with most ISLs. Tree computation costs O(|I| log |S|). FLUTE and OpenMPI implement tree-only aggregation [94, §3.2, §4]. |
| Orbital configuration | 207 · 117 · 12 · 1.52 · 12.4 · 30% | Simulations use 207 Planet satellites, 117 Spire satellites and Planet's 12 ground stations. Reported time-to-accuracy speedup spans 1.52×–12.4×. Accuracy falls sharply at 30% participation [94, §5, §5.4]. |
| Model scope and venue | 100 | The paper is the first in this line to assume ISLs between Earth-observation satellites. Its ~100 Mbps cap, a workshop venue and the absence of a failure model are what a reader must keep beside the speedups. |
| 證據與範圍 | 2026 | OrbitalBrain 是 NINeS 2026 的工作坊論文，作者群包含三位 FedSpace 作者。 |
| 證據與範圍 | 定性範圍 | 它將 FedSpace 的僅地面通訊前提改為星間通訊，讓地球觀測衛星透過星間鏈路彙總聯邦學習的更新。 |
| 證據與範圍 | 4–6 · 7 | 作者估計每天有數百到數千次 ISL 機會，相對地面過頂只有 4–6 次、每次最多 7 分鐘[94, §1]。 |
| 電力與熱條件 | 4 GB · 7.5 W · 7 W | 每顆衛星搭載一片 Jetson Orin Nano 4 GB，功耗約 7.5 W，對上約 7 W 的太陽能輸入。 |
| 證據與範圍 | 100 | ISL 與地面鏈路頻寬的中位數約 100 Mbps，模擬器把 ISL 上限設為 100 Mbps。 |
| 證據與範圍 | 10 s · 0.9 | 超過 10 s 的 ISL 建鏈時間被折算成固定的 0.9 效率係數[94, §2.1, §4, §5.1, App. B]。 |
| 證據與範圍 | 定性範圍 | 更新沿著以 ISL 最多的衛星為根的最短路徑樹彙總，成本為 O(|I| log |S|)，後端是 FLUTE 與 OpenMPI，僅使用樹狀彙總[94, §3.2, §4]。 |
| 證據與範圍 | 207 · 117 · 12 · 1.52 · 12.4 · 30% · 5 · 5.4 | 在模擬的 207 顆 Planet 與 117 顆 Spire 衛星機隊、搭配 Planet 的 12 座地面站上，它報告達標時間加速 1.52 到 12.4 倍，但只有 30% 衛星參與時準確率急遽下降（摘要、§5、§5.4）。 |
| 證據與範圍 | 100 | 這是此研究線中第一篇假設地球觀測衛星之間有 ISL 的論文，而約 100 Mbps 的上限、工作坊的發表場所以及缺少故障模型，是讀者在看加速倍數時必須一併記住的事。 |
1. Where this paper sits1. 這篇論文的位置
OrbitalBrain extends FedSpace’s contact-aware learning with inter-satellite aggregation. Its dated fleet assumptions place it in constellation-edge learning, separate from formation training.
OrbitalBrain 在 FedSpace 接觸感知學習之上加入星間彙總，其附有日期的機隊假設將它定位於星座邊緣學習，與編隊訓練分開。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | 9.3 · 9 | OrbitalBrain closes the federated-learning thread of Section 9.3 and is the most recent of the five service papers in Chapter 9. |
| Evidence and scope | 12 | It descends directly from FedSpace [91], with which it shares Hsieh, Arzani and Chandra. It uses the same Planet fleet and the same 12 ground stations. Thus, the change between the two papers isolates one variable, the existence of inter-satellite links. |
| Earlier communication assumption | 2022 | FedSpace's 2022 premise restricts feasible LEO satellite communication to ground links. |
| Dataset and workload | 100 · 59 | Four years after FedSpace, the group assumes about 100 Mbps median ISLs. Fact-base Part C, item 59 uses this change to emphasize dated assumptions. |
| Capacity and rate | 11 · 10.6 | The paper also matters for Part III of this primer because it is the highest-rate federated-learning design in the group. It is still five orders of magnitude below the per-link rate of the Suncatcher formation in Chapter 11. This explains why Section 10.6 assigns it to a separate regime from orbital training clusters. |
| 證據與範圍 | 9.3 · 9 | OrbitalBrain 為第 9.3 節的聯邦學習主線收尾，也是第 9 章五篇服務論文中最新的一篇。 |
| 證據與範圍 | 12 | 它直接承接 FedSpace[91]，兩者共享 Hsieh、Arzani 與 Chandra 三位作者，並使用同一個 Planet 機隊與同樣 12 座地面站，所以兩篇論文之間的差異只隔離出一個變數：星間鏈路如何存在。 |
| 證據與範圍 | 2022 | FedSpace 在 2022 年寫道衛星間通訊「LEO 衛星目前可行性待驗證」。 |
| 證據與範圍 | 100 · 59 | 四年後同一團隊假設 ISL 的中位數約 100 Mbps，事實庫把這列為這個文獻中必須標上日期之假設最清楚的例子（事實庫 Part C 第 59 項）。 |
| 證據與範圍 | 11 · 10.6 | 這篇論文對本教材第三部分也很重要，因為它是本組中速率最高的聯邦學習設計，卻仍比第 11 章 Suncatcher 編隊每條鏈路的速率低五個數量級，這也是為什麼第 10.6 節把它歸入與軌道訓練叢集相異的區間。 |
2. The problem, measured2. 問題，用量測說話
ISLs provide more frequent contacts than ground passes at the same assumed link rate. Ground communication also consumes more power than local compute in the stated hardware budget.
在相同假設速率下，ISL 比地面過頂提供更頻繁的接觸，而所列硬體預算中，對地通訊的功耗也高於本地運算。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | Qualitative scope | The problem is stated as a count of opportunities. |
| Evidence and scope | 4–6 · 7 | An Earth-observation satellite in the Planet fleet sees a ground station 4–6 times a day for at most 7 minutes each time. Thus, a federated-learning scheme that aggregates only on the ground, as FedSpace does, spends most of each day waiting. In comparison, the same satellite has hundreds to thousands of chances per day to form an ISL with another satellite [94, §1]. |
| Evidence and scope | Qualitative scope | The resource budget is the second half of the problem. |
| Power and thermal conditions | 4 GB · 7.5 W · 7 W · 50 W | A Jetson Orin Nano 4 GB draws about 7.5 W against roughly 7 W solar collection. Ground downlink draws about 50 W. Ground communication therefore costs substantially more power than local compute [94, §2.1, App. B]. |
| Capacity and rate | 100 | The authors take about 100 Mbps as the median bandwidth of both ISLs and ground links. This means the two kinds of link are equally fast and differ only in how often they exist and what they cost to run [94, §2.1, Fig. 1]. |
| 證據與範圍 | 定性範圍 | 問題以「機會的次數」來陳述。 |
| 證據與範圍 | 4–6 · 7 | Planet 機隊中的地球觀測衛星每天看到地面站 4–6 次，每次最多 7 分鐘，所以像 FedSpace 那樣只在地面彙總的聯邦學習方案，每天大部分時間都在等待。 |
| 證據與範圍 | 定性範圍 | 而同一顆衛星每天有數百到數千次機會和另一顆衛星建立 ISL[94, §1]。 |
| 證據與範圍 | 定性範圍 | 資源預算是問題的另一半。 |
| 電力與熱條件 | 4 GB · 7.5 W · 7 W · 50 W | 每顆衛星配備一片 Jetson Orin Nano 4 GB，這塊板子的功耗約 7.5 W，而衛星收集的太陽能約 7 W，地面下行鏈路則耗電約 50 W，所以對地通訊的能耗遠高於機上運算[94, §2.1, App. B]。 |
| 證據與範圍 | 100 | 作者把 ISL 與地面鏈路的頻寬中位數都取為約 100 Mbps，也就是說兩種鏈路一樣快，比較差異在於它們出現的頻率與運作的代價[94, §2.1, Fig. 1]。 |
3. The design3. 設計
The current constellation graph has satellite vertices S and ISL edges I. Model updates aggregate over a shortest-path tree rooted at the highest-degree satellite. Each update follows a shortest path to that root. Recomputing the tree costs O(|I| log |S|) over the current graph [94, §3.2].
Opportunistic links change with satellite motion, favoring cheaply recomputed aggregation structures. Fixed terrestrial ring and all-reduce schedules describe another operating regime. OrbitalBrain uses tree aggregation only [94, §4]. FLUTE supplies federated learning, with OpenMPI communication [94, §4]. Planet's 12 ground stations remain, joined by satellite aggregation sites.
OrbitalBrain 把任一時刻的星座看成一張圖：頂點是衛星 S，邊是當下存在的 ISL I，並在這張圖上彙總模型更新，，彙總位置因此移到軌道上。彙總結構是一棵最短路徑樹，根是當下 ISL 最多的衛星，使根成為連線最好的節點，每顆其他衛星的更新都沿著最短路徑到達它，而這棵樹的計算時間為 O(|I| log |S|)，也就是在當前圖上做一次最短路徑搜尋的成本[94, §3.2]。對這種圖而言，樹是自然的選擇，因為鏈路是機會式的，拓樸隨著衛星進出彼此範圍而改變，所以一個能從現有鏈路便宜地重新計算的結構，比地面訓練叢集在固定網路上使用的環狀或 all-reduce 排程更實用，而論文也確實僅使用樹狀彙總[94, §4]。實作建立在 FLUTE 聯邦學習框架上，以 OpenMPI 作為通訊後端[94, §4]。地面站仍然留在系統裡，因為 Planet 的 12 座地面站是模擬網路的一部分，但衛星也成為可進行彙總的站點。
Original figure note原始圖說出處
Drawn for this primer from Chabra et al., NINeS 2026, §1, §2.1, §3.2, §4, §5.1 and App. B.4. Evaluation setup4. 實驗設定
The simulation uses Planet and Spire fleets and a capped ISL model. A fixed setup-efficiency factor approximates establishment cost, while participation sweeps test one aspect of robustness.
模擬採用 Planet 與 Spire 機隊及設有上限的 ISL 模型，以固定建立效率係數近似建鏈成本，參與率掃描則檢驗部分穩健性。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | Qualitative scope | The evaluation is simulation. |
| Orbital configuration | 207 · 117 · 12 | The constellations are 207 Planet satellites and 117 Spire satellites, with Planet's 12 ground stations, and training runs in FLUTE with an OpenMPI backend [94, §4, §5.1]. |
| Capacity and rate | 100 · 0.9 · 10 s | Each simulated ISL is capped at 100 Mbps, matching the estimated median ISL and ground bandwidth. Setup exceeding 10 s multiplies throughput by 0.9. The model applies this flat efficiency instead of contact-specific acquisition downtime [94, §2.1, §4, §5.1]. |
| Latency comparison | 15 | The ISL topology is the set of opportunities that the orbits produce, an orbit-dependent graph. The propagation latency of individual ISLs remains unspecified in the fact base. In comparison, the Starlink-specific 15-second schedule lies outside these fleets’ scope, which belongs to Starlink rather than to Planet or Spire. |
| Power and thermal conditions | 4 GB · 7.5 W · 7 W | Each satellite's compute is a Jetson Orin Nano 4 GB at about 7.5 W against about 7 W of solar input [94, App. B]. |
| Participation robustness | 30% | Failure modeling remains open, and the robustness experiment varies the fraction of satellites participating, with accuracy falling sharply at 30% participation [94, §5.4]. |
| 證據與範圍 | 定性範圍 | 評估是模擬。 |
| 證據與範圍 | 207 · 117 · 12 | 星座是 207 顆 Planet 衛星與 117 顆 Spire 衛星，搭配 Planet 的 12 座地面站，訓練在 FLUTE 中以 OpenMPI 後端執行[94, §4, §5.1]。 |
| 證據與範圍 | 100 | 在鏈路方面，模擬器把每條 ISL 的上限設為 100 Mbps，與論文估計的 ISL 與地面鏈路頻寬中位數約 100 Mbps 一致。 |
| 證據與範圍 | 10 s · 0.9 | 對雷射或射頻鏈路的建立，它的處理方式是只要建鏈超過 10 s 就把吞吐量乘以 0.9 的效率係數，，每次接觸的擷取時間成本則列於模型之外[94, §2.1, §4, §5.1]。 |
| 證據與範圍 | 定性範圍 | ISL 拓樸是軌道產生的機會集合，隨軌道變化的圖。 |
| 證據與範圍 | 定性範圍 | 各自 ISL 的傳播延遲待補充。 |
| 證據與範圍 | 15 | 15 秒排程是 Starlink 的特性，列於 Planet 與 Spire 的範圍之外。 |
| 電力與熱條件 | 7.5 W · 4 GB · 7 W | 每顆衛星的運算是功耗約 7.5 W 的 Jetson Orin Nano 4 GB，對上約 7 W 的太陽能輸入[94, App. B]。 |
| 證據與範圍 | 30% | 故障模型待建立，穩健性實驗改變的是參與衛星的比例，在 30% 參與率時準確率急遽下降[94, §5.4]。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Time-to-accuracy speedups of 1.52× to 12.4× [94, §5].達標時間加速 1.52 到 12.4 倍（摘要、§5）。 | Simulation on 207 Planet and 117 Spire satellites with 12 Planet ground stations. ISLs capped at 100 Mbps. The range spans constellations, datasets and baselines.在 207 顆 Planet 與 117 顆 Spire 衛星、12 座 Planet 地面站上模擬。ISL 上限 100 Mbps。範圍涵蓋相異星座、資料集與基準。 |
| Hundreds to thousands of ISL opportunities per day against 4–6 ground passes of at most 7 min [94, §1].每天數百到數千次 ISL 機會，相對 4–6 次、每次最多 7 分鐘的地面過頂[94, §1]。 | Derived from orbital geometry of the two fleets, with flown-ISL verification pending.由兩個機隊的軌道幾何推導，，飛行 ISL 硬體的量測則待補充。 |
| Median ISL and ground-link bandwidth about 100 Mbps [94, §2.1, Fig. 1].ISL 與地面鏈路頻寬中位數約 100 Mbps[94, §2.1, Fig. 1]。 | An estimate used to set the simulator's cap. The fact base records flown Planet or Spire ISL verification pending.用來設定模擬器上限的估計值。事實庫將已飛行 Planet 或 Spire ISL 硬體列為待查證。 |
| Accuracy drops sharply at 30% participation [94, §5.4].參與率 30% 時準確率急遽下降[94, §5.4]。 | The only robustness result. Satellite or link failures remain outside the explicit model.唯一的穩健性結果。衛星與鏈路故障模型仍待補充。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| How many bytes can one 7-minute pass carry at 100 Mbps?以 100 Mbps 計，一次 7 分鐘過頂能傳多少位元組？ | \( 100\,\text{Mb/s} \times 420\,\text{s} / 8 \) | 5.25 GBcomputed. 100 Mbps from §2.1, ≤7 min from §1計算值。100 Mbps 來自 §2.1，≤7 min 來自 §1 |
| How long does a ~45 MB ResNet-18 take over one ISL?約 45 MB 的 ResNet-18 經一條 ISL 要傳多久？ | \( 45 \times 8\,\text{Mb} / 100\,\text{Mb/s} \) | 3.6 scomputed. Model size as in the fact base, ignoring setup計算值。模型大小依事實庫，忽略建鏈時間 |
| What share of a day is ground contact?地面接觸占一天的多少？ | \( (4 \ldots 6) \times 7\,\text{min} / 1{,}440\,\text{min} \) | ≈ 1.9–2.9%computed from §1, taking every pass at its 7-min maximum由 §1 計算，每次過頂都取 7 分鐘上限 |
| On what contact length does a 10 s setup cost exactly 10%?在多長的接觸上，10 s 建鏈恰好損失 10%？ | \( 10\,\text{s} / (1 - 0.9) \) | 100 sCalculated using the upper 30 s setup in the modern-terminal range of 2–30 s (Part C, item 18). A 300 s contact would retain the same factor.計算值。30 s 的建鏈（現代終端 2–30 s 範圍的上限，Part C 第 18 項）需要 300 s 的接觸才能對應同一係數 |
| How far below a Suncatcher link is the ISL cap?ISL 上限比 Suncatcher 每條鏈路低多少？ | \( 10\,\text{Tb/s} / 100\,\text{Mb/s} \) | 10⁵, five orders of magnitudecomputed. On the order of 10 Tbps per link from [36]計算值。每條鏈路約 10 Tbps 量級來自 [36] |
| How much more power does the downlink draw than the compute?下行鏈路比運算多耗多少電？ | \( 50\,\text{W} / 7.5\,\text{W} \) | ≈ 6.7×computed. Both from §2.1 and App. B計算值。兩者皆來自 §2.1 與 App. B |
6. Limits and open questions6. 範圍與待解問題
- A workshop paper. NINeS 2026 is a refereed OASIcs workshop. The compact peer-reviewed design leaves per-ISL latency, contact duration and per-dataset convergence curves unspecified in the fact base.工作坊論文。NINeS 2026 是有審查、收錄於 OASIcs 的工作坊，所以設計經過同儕審查但篇幅精簡。系統讀者會想知道的幾個量，例如每條 ISL 的延遲、接觸時長與各資料集的收斂曲線，在事實庫中都待補充。
- A ~100 Mbps ceiling. The Planet-class estimate is three orders below Starlink's claimed 100 Gbps sustained per laser company. It is five orders below Suncatcher's roughly 10 Tbps target. Results therefore describe constellation-edge learning, while data-center training needs another regime (Part C, items 14 and 59).約 100 Mbps 的天花板。這個上限是論文自己對 Planet 等級衛星的估計，比 Starlink 宣稱的每具雷射 100 Gbps 持續速率 公司 低三個數量級，比 Suncatcher 每條鏈路約 10 Tbps 量級的目標低五個數量級，所以結果描述的是星座邊緣學習，軌道資料中心訓練則屬另一種體制（Part C 第 14、59 項）。
- Setup as a flat 0.9. Folding link establishment into a constant throughput factor is accurate only for contacts of about 100 s with a 10 s setup. For short opportunistic contacts or 30 s acquisitions the real loss could be much larger. Every topology and routing paper except a few treats setup as free (Part C, items 18 and 56).建鏈被當成固定的 0.9。把鏈路建立折算成固定的吞吐量係數，只有在 10 s 建鏈、約 100 s 接觸時才準確。對短暫的機會式接觸或 30 s 的獲取，真正的損失可能大得多，而除了少數幾篇之外，所有拓樸與路由論文都把建鏈當成免費（Part C 第 18、56 項）。
- Failure modeling remains open. Satellites and links remain operational in simulation. Accuracy collapses at 30% participation. This suggests opportunistic graph partitions could impair learning substantially.故障模型待建立。模擬中衛星與鏈路保持運作，而參與率掃描顯示準確率在 30% 時崩落，這暗示 ISL 圖的分割（在機會式鏈路下很可能發生）會造成嚴重傷害。
- Power is tight. A 7.5 W board on about 7 W of solar input requires a duty cycle or supplementary energy to train. The paper's energy figures come from an appendix rather than from flight. Chapter 8 records that the only in-orbit COTS thermal measurement throttled within about 50 minutes.電力很緊。7.5 W 的板子對上約 7 W 的太陽能輸入，需要工作週期或額外能源才能訓練，而論文的能源數字來自附錄，飛行資料則待補充。第 8 章記載唯一的在軌 COTS 熱量測在約 50 分鐘內就進入降頻。
7. How it connects7. 與其他內容的關聯
- Section 9.3: OrbitalBrain is the ISL end of the federated-learning table, and Figure 9.3 draws its ISL opportunities against 4–6 ground passes.第 9.3 節：OrbitalBrain 是聯邦學習表中有 ISL 的那一端，圖 9.3 畫出它的 ISL 機會與每天 4–6 次地面過頂的對比。
- Section 9.5: at about 100 Mbps it is the highest-rate FL paper on the link-rate ladder and still below every CDN and inference assumption.第 9.5 節：約 100 Mbps 讓它成為鏈路速率之梯上速率最高的 FL 論文，卻仍低於所有 CDN 與推論的假設。
- Section 9.1, Section 9.2 and Section 9.4 cover neighboring CDN and inference services. Among these, on-board inference has flight evidence.第 9.1 節、第 9.2 節與第 9.4 節：與它並列的 CDN 與推論服務，其中只有機上推論真正飛過。
- Section 3.3: the ground stations that remain in OrbitalBrain's network and share the aggregation role with satellites.第 3.3 節：仍留在 OrbitalBrain 網路中、與衛星共同擔任彙總點的地面站。
- Section 10.6 compares three regimes. OrbitalBrain anchors constellation-wide learning over hundreds of kilometers at Gbps-or-less rates. Chapter 11 supplies the Tbps formation-cluster comparison.第 10.6 節：三種區間，OrbitalBrain 是星座尺度區間（跨數百公里、Gbps 或更低的鏈路）的代表，對照第 11 章的 Tbps 編隊叢集。
- FedSpace [91] is its ground-only predecessor on the same fleet. FedSat [90] is the root of the line. StarCDN is the other service that routes over ISLs, at 100 Gbps. Suncatcher [36] and the laser setup-delay study mark the link rate and acquisition time that this design abstracts away.FedSpace[91]是它在同一機隊上僅用地面鏈路的前作。FedSat[90]是這條研究線的起點。StarCDN 是另一個經由 ISL 路由的服務，速率 100 Gbps。Suncatcher[36]與雷射建鏈延遲研究標出了本設計抽象掉的鏈路速率與獲取時間。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 1.52–12.4× | Time-to-accuracy speedup, simulated [94, §5] | 達標時間加速倍數，模擬（摘要、§5） | |
| ~100 Mbps | Median ISL and ground-link bandwidth. Simulator ISL cap [94, §2.1, §4] | ISL 與地面鏈路頻寬中位數。模擬器的 ISL 上限[94, §2.1, §4] | |
| >10 s → 0.9 | ISL setup time folded into a throughput efficiency factor [94, §5.1] | ISL 建鏈時間折算成的吞吐量效率係數[94, §5.1] | |
| 100s–1,000s vs 4–6 | Daily ISL opportunities vs ground passes of ≤ 7 min [94, §1] | 每天的 ISL 機會對 ≤ 7 min 的地面過頂次數[94, §1] | |
| ≈7.5 W · ≈7 W · ≈50 W | Jetson Orin Nano 4 GB, solar input, ground downlink [94, §2.1, App. B] | Jetson Orin Nano 4 GB、太陽能輸入、地面下行[94, §2.1, App. B] | |
| 207 · 117 · 12 | Planet satellites, Spire satellites, Planet ground stations [94, §5.1] | Planet 衛星、Spire 衛星、Planet 地面站[94, §5.1] | |
| O(|I| log |S|) | Cost of the shortest-path aggregation tree. Tree aggregation only [94, §3.2, §4] | 最短路徑彙總樹的成本。採樹狀彙總[94, §3.2, §4] | |
| 30% | Participation at which accuracy drops sharply [94, §5.4] | 準確率急遽下降時的參與率[94, §5.4] |