Embedded Files
Spacecraft Compartmentalized Autonomous Learning and Edge-Computing System
Mission of the project
Mission of the project
The Spacecraft Compartmentalized Autonomous Learning and Edge-Computing System (SCALES) represents a groundbreaking advancement in on-orbit machine learning (ML) and edge computing. This integrated hardware and software system leverages Commercial Off-The-Shelf (COTS) edge computing technology and NASA JPL's F' (F Prime) software framework to create a robust and reusable environment for ML algorithm execution and training directly on spacecraft.
Primary Compute System
Primary Compute System
The SCALES system shall be able to serve as the primary compute system for small satellites, allowing use of external sensors.
AI and ML Applications
AI and ML Applications
The SCALES system shall be able to reliably deploy and retrain AI and ML models in LEO environments.
F Prime Flight Software
F Prime Flight Software
The SCALES system will work in partnership with JPL to integrate their F Prime flight software.
Commercial Off-The-Shelf
Commercial Off-The-Shelf
The SCALES system shall integrate COTS hardware to make the system for accessible and user-friendly.
Open Source
Open Source
The SCALES system shall open source upon completion, contributing to the accessibility and ease of use of the system.
Easily Adaptable
Easily Adaptable
The SCALES system shall allow for multiple types of AI / ML models to be used, allowing for a variety of different use cases.
Technology Overview
Technology Overview
We propose an integrated hardware / software product that allows for direct deployment of advanced ML algorithms onboard autonomous space systems. the SCALES system combines COTS edge computing hardware with the resilient and reusable F Prime software framework from NASA JPL to create a single dynamic system. ML algorithms can be deployed of autonomous tasks, but the on-board compute can also be used for in-situ tuning of the algorithm using a reinforcement learning system.
This component would occupy an approximately 1U form factor and be designed as a stand-alone module that can be attached to larger spacecraft in either CubeSat or ESPA-class SmallSat scales.
Overall Architecture
Overall Architecture
Potential Impact
Potential Impact
A reusable hardware / software product for deploying AI / ML on spacecraft would rapidly accelerate the pace of deploying complex SmallSat systems by reducing development and integration times.
Autonomous detection networks can analyze, task, and generate insights on transient phenomena without a human decision maker in the loop. Greatly improving response times and collecting data that would otherwise be lost.
Satellite control and sensing systems can be dynamically trained and respons to new datasets or faults without the need for human intervention.
Pictured: Integration of NVIDIA Jetson Nano in flight stack on the BroncoSat-1, 1.5U CubeSat
Technology Development Objectives
Technology Development Objectives
Primary Objective: Qualify a flight ready and reusable hardware / software product for the deployment of AI / ML algorithms on Small Spacecraft.
Extend current F Prime capabilities to support edge compute hardware and GPU acceleration
Starting TRL: 4 | Ending TRL: 6
Primary MOP: Inference Time (YOLO Algorithm)
SoA: 4 Images/Sec (using Mars Ingenuity Helicopter Processor)
Target: ~1,400 Images/Sec (using SCALES dedicated co-processor)
Secondary KPP: Onboard retraining of a computer visions algorithm with a mean average precision (mAP) > 50%.
Complete an Environmental Testing Campaign for a compact compute stack with COTS components according to NASA GEVS-7000B
Providing higher levels of on-board autonomy for Small Spacecraft Systems exploring beyond Earth Orbit enables responsive science collection and the potential for close to real time on-board decision making.
Technology End Users
Technology End Users
Satellite Developers: Fault tolerant satellite control through anomaly detection and correction.
Earth Science: Enabling autonomous detection of transient phenomena and re-tasking of on-orbit assets.
Exploration Missions: Allowing for autonomous mission planning an tasking without the need for constant human oversight.
Questions?
Questions?
Contact broncospacelab@gmail.com to get more information on the project.
Page updated
Google Sites
Report abuse