OrbitNet, a compact machine-learning model, posted the best reported result in the Starlink orbit-prediction test. Its mean absolute error—the average size of the position miss—was 3.34 metres, while its root mean squared error, a score that gives greater weight to larger misses, was 22.80 metres. The model had 78,432 parameters and a reported 48.6% lower mean absolute error than the second-best result.
In six zero-shot tests, OrbitNet had the best mean absolute error on every dataset. Its root mean squared error was best on ICEYE, LEMUR and SKYSAT, second on CAPELLA, and lower for DLinear or DASR on ASTROCAST or KINEIS, respectively.
What the tests measured
The study examines whether injecting velocity-aware motion cues into position representations improves satellite orbit prediction compared with simply appending velocity channels. OrbitNet pairs that refinement with orbital segment modeling and a lightweight prediction head.
To build the data, the researchers used publicly accessible Space-Track two-line element (TLE) records and Orekit, converting them into uniformly sampled one-minute sequences containing position and velocity variables. Starlink accounted for 260 satellites, with 208 used for training and 52 for testing.
Each sample used a 512-step history to predict a 90-step future position sequence from six state variables. The researchers calculated mean absolute error and root mean squared error after denormalising the data, reported both in metres, and averaged them over the three spatial coordinates; comparison models used the same six-dimensional input and three-dimensional position target.
Performance beyond Starlink
On the six unseen datasets, OrbitNet’s reported mean absolute error/root mean squared error pairs ranged from 4.55/10.00 metres on SKYSAT to 32.68/106.97 metres on KINEIS. The other results were 8.33/27.46 metres on ASTROCAST, 14.37/19.75 on CAPELLA, 14.12/38.47 on ICEYE and 10.34/38.82 on LEMUR.
Across the six datasets, OrbitNet’s average mean absolute error was 14.07 metres, versus 45.47 metres for DASR and 193.20 metres for KiGRU. The reported mean absolute error gains over second-best baselines were 59.2% on SKYSAT and 32.5% on ASTROCAST.
Which parts mattered
On Starlink, the ablation comparisons reported mean absolute error/root mean squared error pairs of 14.48/31.94 metres for the baseline, 5.43/29.83 with velocity refinement, 4.27/27.39 with segment modeling, and 3.34/22.80 for full OrbitNet.
Other refinement tests reported 91.56/136.93 metres for a correlation-only variant, 8.73/34.06 for simple summation and 3.34/22.80 for coupling. A learnable fusion coefficient reported 3.34/22.80, compared with 4.29/26.27 when the coefficient was fixed at 1.00; removing positional embedding reported 6.95/30.50, compared with 3.34/22.80 for the learnable version.
The learned dependency matrix included cross-links between position and velocity variables. Examples included x receiving weight 0.187 from v_x, z receiving 0.186 from v_x, x receiving 0.152 from v_y, z receiving 0.220 from x, and v_z receiving 0.162 from v_y.
Errors grew over longer forecasts
At 180, 360 and 720 forecast steps, OrbitNet’s mean absolute error/root mean squared error pairs were 3.78/25.30, 6.30/27.93 and 10.81/31.94 metres, respectively. The paper reported OrbitNet as the best among the selected baselines at all three horizons and attributed the larger errors at longer horizons generally to error accumulation.
Cautions around the numbers
The Starlink comparison reports no confidence intervals or repeated-run variability, while the six zero-shot results include no uncertainty estimates.
The record is an arXiv preprint, arXiv:2608.23728v1, dated 24 August 2026; its front matter says the manuscript was submitted to ACM.
Paper data and sources
Original title: Velocity-coupled Representation Refinement for Satellite Orbit Prediction
Authors: Yue Yang, Zhiqiang Wu, Saiyu Qi, Fan Ma
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text