Executing vision-language models (VLMs) aboard low-Earth orbit satellites enables flexible processing of observation data without first downlinking it to the ground. On commercial off-the-shelf payloads, however, a request may span multiple execution windows: power and thermal constraints, fault-triggered reboots, and competing payload demands can interrupt inference before it completes. Each shutdown discards volatile activations and KV caches, forcing an unprotected request to repeat work after restart. We present OrbitWeaver, a framework that preserves VLM inference progress across onboard interruptions through coordinated checkpointing and recovery. Its phase-aware checkpointing selects recovery boundaries according to the state lifetimes and computational characteristics of vision encoding, language-model prefill, and decoding, balancing storage traffic against recomputation. Its dependency-aware dynamic recovery schedules restoration from partially completed states across the layer and token-chunk dimensions. By overlapping checkpoint and weight reads with recomputation, it reduces restart latency; the recovery planner computes an optimal execution order under its scheduling model within 100 ms. Experiments on satellite payload hardware using on-orbit telemetry traces show up to a 4.07-fold end-to-end speedup and a 3.56-fold reduction in SSD write volume over the evaluated baselines. A 72-hour deployment on two low-Earth orbit satellites demonstrates that OrbitWeaver retains inference progress and resumes execution across repeated bus-enforced shutdowns.
Figures & tables
Figure 1. Satellite architecture diagram. TT&C = Telemetry, Tracking, and Command; OBC = On-Board Computer; FDIR = Fault Detection, Isolation, and Recovery; EPS = Electrical Power System. Satellite computing architecture with payload sensing and computing components connected to bus support and control subsystems.
Figure 2. The temperature trace measures payload enclosure temperature. TT&C = Telemetry, Tracking, and Command; OBC = On-Board Computer; FDIR = Fault Detection, Isolation, and Recovery; EPS = Electrical Power System. Three-panel workflow showing the payload enclosure temperature trace, thermal-triggered payload power off, spacecraft bus recovery control, and VLM inference stages from image tokens to KV cache.
Figure 3. Overview Satellite computing architecture with payload sensing and computing components connected to bus support and control subsystems.
Figure 4. OrbitWeaver’s checkpointing strategies. Comparison of layer-wise, chunk-wise, and block-wise prefill checkpointing, with computation blocks, intermediate states, and checkpoint boundaries.
Figure 5. Storage–recomputation trade-off of different checkpointing granularities. Trade-off between checkpointing overhead and lost computation for different block sizes.
Figure 6. Checkpoint storage layout for prefill, vision, and decode. Storage layout for persisted checkpoints across prefill, vision, and decoding, including KV cache and hidden-state buffers.
Figure 7. Impact of recovery latency on interrupted VLM inference with a 64K-token request. Two line plots showing that end-to-end time increases with recovery time on AGX and Orin. AGX reports Qwen3-VL 2B, 4B, and 8B, while Orin reports Qwen3-VL 2B and 4B.
Figure 8. Fixed recovery orders expose avoidable I/O stalls. Four examples of static and optimal recovery orders over two layers and three chunks. Orange boxes denote model parameters, dark green boxes denote KV checkpoints, and arrows denote compute order.
Figure 9. Frontier Dependency. (a) A concrete staircase frontier after a shutdown, with f1 , f2 , f3 , and f4 marking the completed prefix of each layer. (b) Dependency examples for three compute tasks, C4,4 , C4,5 , and C4,6 . Frontier dependency diagram. Panel (a) shows a staircase frontier with markers for f1, f2, f3, and f4. Panel (b) shows dependency examples for three compute tasks in the fourth layer.
Task
Symbol
Type
Execution time
Load model weights
Wi
I/O
τW
Load hidden states
Hi,j
I/O
τH
Load KV cache
Ki,j
I/O
τK
Execute block
Ci,j
Compute
τC
Table 1. Tasks in the dynamic recovery scheduling model.
Figure 10. End-to-end performance under different setups Six bar charts compare interrupted end-to-end time across checkpointing baselines and OrbitWeaver for different interruption traces, devices, and VLMs.
Figure 11. Recovery latency under different setups. Six bar charts compare recovery time across checkpointing baselines and OrbitWeaver at different completed-progress points, devices, and VLMs.
Trace
A
B
C
D
E
Satellite
Sat-1
Sat-1
Sat-2
Sat-2
Dellingr ( Hatfield, 2018 )
Mean window (s)
36.4
42.3
121.7
108.2
63.0
Table 2. Interruption traces used in evaluation.
Figure 12. Inference latency under different setups when no shutdowns occur. Six bar charts compare uninterrupted inference time across checkpointing baselines and OrbitWeaver at different devices, models, and prompt lengths.
Figure 15
Figure 15. An example in-orbit VLM execution across interrupted windows on Sat-1. Temperature refers to the payload enclosure temperature. Two payload enclosure temperature and power traces compare OrbitWeaver and no-checkpoint execution across three interrupted execution windows. OrbitWeaver resumes after shutdowns and reaches decoding, while no checkpoint restarts from the beginning after each shutdown and fails to complete.
Onboard vision-language models could enable satellites to answer queries directly, but exhaustive tiled inference over high-resolution imagery is slow and energy-intensive. We identify answer-invariant token redundancy (AITR): image tiles and vision tokens that can be removed without changing the final answer. We present Rift, a two-stage system that performs query-conditioned tile pruning followed by elastic prefill to reduce token budget. We evaluate it on LLaVA-1.5 7B running on Jetson AGX Orin. Compared with exhaustive tiled inference, Rift reduces energy by 78% and latency by 69%, while increasing accuracy from 45% to 73%.
As Earth Observation data generation outpaces downlink bandwidth and human-in-the-loop processing, a widening gap has emerged between onboard collection and actionable ground intelligence. This paper presents NAVI-Orbital, a software system deployed on a Low Earth Orbit (LEO) spacecraft. On April 16, 2026, NAVI-Orbital achieved what is, to the authors' knowledge, the first in-orbit demonstration of a vision-language model performing autonomous multi-modal inference entirely onboard. NAVI-Orbital uses a local vision-language model (Gemma 3) to classify each captured scene, produce a text description of its content and the relationships between its features, and respond to operator follow-up via natural-language dialogue. The system is re-tasked through plain-English prompts in place of conventional command sequences, and is orchestrated by a graph-based state machine (LangGraph) coordinating dedicated agents for detection and dialogue. Results across ground benchmarking (88.16% accuracy on the 7,960-image curated AID benchmark), Flatsat validation, and live in-orbit captures of newly acquired, previously unseen Earth imagery (including uncorrected YAM-9 imagery, processed onboard with hardware-accelerated GPU inference and no fine-tuning for the flight instrument) demonstrate the feasibility of running foundation models on satellite-class edge computers to invert the conventional acquire-then-downlink-everything bandwidth profile through semantic compression of Earth observations in-orbit.
Juan Manuel Delfa Victoria, Taran Cyriac John, Andrew W. Herson
Jet Propulsion Laboratory California Institute of Technology Pasadena, CA, USA · School of Engineering and Computer Science Victoria University of Wellington Wellington 6140, New Zealand · Loft Orbital 321 11th St, San Francisco, CA 94103
Deploying vision-language models (VLMs) in resource-constrained settings demands low latency and high throughput, yet existing compact VLMs often fall short of the inference speedups their smaller parameter counts suggest. To explain this discrepancy, we conduct an empirical end-to-end efficiency analysis and systematically profile inference to identify the dominant bottlenecks. Based on these findings, we develop optimization recipes tailored to compact VLMs that substantially reduce latency while preserving accuracy. These techniques cut time to first token (TTFT) by 53% on InternVL3-2B and by 93% on SmolVLM-256M. Our recipes are broadly applicable across both VLM architectures and common serving frameworks, providing practical guidance for building efficient VLM systems. Beyond efficiency, we study how to extend compact VLMs with structured perception outputs and introduce the resulting model family, ArgusVLM. Across diverse benchmarks, ArgusVLM achieves strong performance while maintaining a compact and efficient design.