Pose Bowl: Spacecraft Pose Detection and Estimation

Determining the pose of a generic upper stage specifically is an essential capability for enabling in space inspection since the geometry of the target vehicle is not always known beforehand. This challenge seeks to develop an algorithm that can generically determine the relative pose of a target vehicle and improve that estimate through multiple observations.

Inspector spacecraft are designed to conduct independent, low-cost, in-space inspections of other ships and may deploy off a host spacecraft to examine its exterior for damage. Inspector spacecraft typically have limited computing resources, but complex computing demands. Minimally, they must be able to locate a host ship and maintain an understanding of its position and orientation, or spatial pose. A successful proof of concept for a low-cost inspector spacecraft was demonstrated by NASA's Seeker. For this proof of concept, Seeker's object detection algorithm was trained on a specific type of spacecraft. A more useful version of this software would be generalized so that it could detect any type of spacecraft. It is also essential that future versions are able to determine the pose of the target spacecraft across a sequence of images.

In this challenge, solvers will help NASA develop algorithms that could be run on inspector spacecraft. The competition was held in two tracks, with different associated prizes. In the Detection Track, solvers develop object detection solutions that identify a bounding box around a spacecraft in an image. In the Pose Estimation Track, solvers develop solutions that identify the position and orientation (pose) of spacecraft across sequences of images.

The R5 CubeSat program designed the Pose Bowl competition to produce an ML model that could bound an unknown, uncooperative spacecraft target in an image to produce a bearing measurement. We wanted to know ‘can the competition format produce a flyable product with minimal changes’? 

We specifically set up the competition requirements to produce a flyable product that can run in the loop with our GN&C system in realtime. 

The 2nd place winner of the competition provided the fastest solution. A highly skilled intern took that model, optimized it further for our flight computer, and then integrated it into a Core Flight Software application. No changes were made to the model weights. We only changed the resolution of input images and configured the framework for calling the model. 

We have conducted Hardware In The Loop tests with our camera as well as testing it on real in space imagery of a spacecraft it was not trained on. Both tests succeeded and ran under the requirement of once per second. I believe this demonstrates the answer to the question ‘Can the competition format produce a flyable product with minimal changes’ is yes.

Just in support of this challenge, the project owner needed to generate a lot of useful data that will be used in the future.

This is a mature technology (TRL 9) — the realistic path in is usually NASA's Technology Transfer Program: licensing an existing NASA patent, or a Space Act Agreement to use NASA facilities/expertise directly. NASA also runs a startup licensing program with no upfront fee for companies formed to commercialize a specific NASA technology.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.