Software Engineer I, Perception (New Grad)

True Anomaly•Long Beach, CA

2d•$80,000•Onsite

About The Position

Space is a warfighting domain. True Anomaly seeks those with the talent and ambition to build the technology that secures it. True Anomaly delivers decisive capabilities for space superiority. We build autonomous spacecraft, advanced payloads, mission software, and space-based interceptors — enabling the U.S. and its Allies to secure the space environment and counter threats from the ultimate high ground. You'll work on hybrid perception systems combining classical computer vision with modern deep learning for autonomous spacecraft: building multi-object tracking pipelines that fuse neural network detections with Kalman filtering, developing coordinate transformation chains from pixels to orbital frames, training models on synthetic space imagery, and deploying algorithms onboard under strict compute/power constraints. Your work enables spacecraft to detect objects against star fields, track multiple targets through occlusions, discriminate threats from decoys, and generate angle measurements for navigation — using both classical geometric methods and learned representations where each approach excels. This is entry-level work blending traditional robotics perception with modern ML. You'll implement Extended Kalman Filters, train neural networks in PyTorch, write C++ flight code, and see your algorithms operate in orbit. This is a 3 month temporary employment engagement. There is potential to convert to regular employment based on performance and business need.

Requirements

Currently pursuing or recently completed Bachelor's or Master's degree in computer science, electrical engineering, robotics, aerospace engineering, or related technical field
Coursework in both computer vision and estimation theory (or willingness to learn both)
Proficiency in Python; some exposure to C++ (we'll teach you more)
Familiarity with either classical tracking (Kalman filters, data association) OR deep learning (PyTorch, training neural networks)
Understanding of linear algebra, probability, and coordinate transformations
Ability to read research papers from both robotics (ICRA, IROS) and ML venues (CVPR, NeurIPS) and implement algorithms
Strong debugging skills: tracking down lost tracks, numerical instability, and model failure modes
Eagerness to learn the intersection of classical perception and modern ML
U.S. Citizen (required for facility access and government contracts)

Nice To Haves

Experience with Extended Kalman Filters, multi-object tracking, or state estimation
Familiarity with neural network training in PyTorch or TensorFlow: object detection (YOLO, Faster R-CNN), classification, or segmentation
Understanding of coordinate frames: camera intrinsics/extrinsics, quaternions, rotation matrices, ECI/LVLH/RIC frames
Exposure to model optimization for edge deployment: quantization (INT8, FP16), ONNX, TensorRT
Experience with OpenCV, image processing pipelines, or classical feature extraction
Coursework in optimal estimation, sensor fusion, or probabilistic robotics (Kalman/particle filters)
Prior work with synthetic data generation, Blender/Unreal for rendering, or domain randomization
Understanding of data association algorithms: Hungarian algorithm, auction algorithm, JPDA
Familiarity with tracking-by-detection pipelines: detection → association → update → track management
Experience debugging visual systems: false positives, missed detections, track ID switches, covariance tuning
Prior internship or project deploying algorithms to embedded systems (Jetson, mobile, ROS)
Exposure to multi-modal perception: fusing camera + IMU, camera + lidar, or learned sensor fusion

Responsibilities

Implement classical tracking algorithms: Extended Kalman Filters for state estimation, Hungarian algorithm for data association, track management logic (tentative/confirmed/coasted tracks)
Train neural networks for detection and classification: YOLO for object detection, ResNet-based classifiers for threat discrimination, learned appearance features for re-identification
Build hybrid perception pipelines: neural network detections → classical tracking → coordinate transformations → angle-only measurements for navigation
Develop image processing chains: hot pixel filtering, adaptive thresholding, centroiding, connected component analysis, star catalog matching
Deploy models to edge hardware: quantize neural networks (INT8), integrate with C++ inference engines (ONNX Runtime, TensorRT), optimize for space-qualified processors
Implement coordinate transformations: pixel → camera frame → body frame → Earth-Centered Inertial (ECI), accounting for lens distortion and attitude uncertainty
Generate synthetic training data: render spacecraft in Blender with domain randomization (lighting, attitudes, backgrounds), create labeled datasets for rare scenarios
Validate end-to-end performance: software-in-the-loop simulation, processor-in-the-loop testing, hardware-in-the-loop with real camera feeds