PUMA-Former paper proposes uncertainty-aware offloading for MEO-LEO satellite edge computing
A new arXiv paper describes a framework that lets a MEO coordinator offload workloads to LEO satellites despite stale or missing CPU reports, trading higher energy for service gains.
In plain words
In satellite networks, workloads are handed to moving processors whose available computing power is reported only periodically, so decisions rely on outdated information. The paper proposes PUMA-Former, a method where a medium-Earth-orbit coordinator estimates uncertain satellite capacity and allocates work accordingly. In the authors' simulations it completed more work and cut modeled delays, but used more energy than learning baselines.
What was announced
The framework addresses a coordination problem specific to medium Earth orbit and low Earth orbit satellite networks: periodic status reports leave CPU states stale or missing, so offloading decisions are sensitive to uncertain capacity and congestion at shared destinations. PUMA-Former, described in a cross-listed arXiv abstract, has a MEO coordinator that maintains memory of distinct reports and links available evidence, capacity estimates, allocation commands, and execution feedback. Latent-delay attention weighs plausible measurement-age hypotheses, while historical innovations refine CPU estimates beyond the latest received report, and queue-error-aware fusion incorporates auxiliary pressure while accounting for backlog uncertainty.
Those inferred resource estimates feed a structured Transformer allocator that makes joint decisions on destination, partition, bandwidth, CPU, and power, with bounded risk-aware calibration adjusting workload splits for uncertainty and shared destination load. In a traffic-driven development evaluation, the authors report an 8.62% increase in admitted compute-work completion, a 34.78% reduction in mean modeled service delay, and a 71.08% reduction in mean episode-P95 delay relative to ATLA-adapt, plus improvements over fixed-half round-robin offloading in completion and normalized energy. The abstract notes that service gains over the learning baselines come with higher normalized energy, and all figures are from the authors' own evaluation rather than independent testing.
What it does not show
- Evidence is limited to the paper's abstract; the full article was not reviewed, so method details and evaluation setup cannot be verified.
- All reported performance figures come from the authors' own traffic-driven development evaluation and have not been independently confirmed.
- The abstract does not state whether PUMA-Former has been deployed or validated on operational satellite hardware.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.