Research ยท Benchmarks

Benchmarks

Held-out evaluation suites for orbital ML and distributed compute. Train on the open datasets, submit predictions for the held-out test inputs, and compare against baselines - we evaluate results, not methods.

Orbital Intelligence

OrbML-Predict
Orbit determination from sparse observations.
Held-out
ConjunctionNet
Conjunction probability estimation.
Held-out
ManeuverDetect
Maneuver detection and timing.
Held-out
SatClass
Satellite classification.
Held-out
ReentryPredict
Re-entry time prediction.
Held-out

Distributed Compute

FedSpace
Federated learning across orbital nodes.
Held-out
PartitionBench
Model partitioning under node heterogeneity.
Held-out
SyncSchedule
Synchronization under intermittent links.
Held-out
CompressionBench
Gradient compression efficiency.
Held-out

How to participate.

Get the training and validation sets from the data portal, use any approach - ML, physics-based, or hybrid - then upload predictions for the test inputs. We score against held-out labels and return results.