Summary
What you’ll impact
Our company is seeking a Machine Learning Engineer specializing in perception to develop and deploy 3D perception systems for autonomous construction equipment. The role involves designing early-fusion architectures, building robust perception pipelines, and optimizing models for edge deployment on embedded hardware.
Responsibilities
What you'll do
- Design Early Fusion Architectures: Develop and train state-of-the-art models (e.g., BEV-based transformers) that fuse raw Lidar and Camera data to solve for object detection and semantic segmentation.
- Tackle "Messy" Physics: Build perception systems robust enough to handle dynamic occlusion (seeing the robot’s own arm/bucket), particulates (dust, snow, rain), and high-vibration conditions.
- Deploy to the Edge: Optimize models for inference on embedded hardware. You will debug system-level issues, such as sensor calibration drift and latency bottlenecks.
- Collaborating with other teams to create state-of-the-art representations for downstream use cases.
Requirements
What you’ll bring
- Production ML Experience: 5+ years of experience taking deep learning models from research to real-world production using PyTorch, Tensorflow, or JAX.
- 3D Geometry & Calibration: You have a deep understanding of SE(3) transformations, homogeneous coordinates, and intrinsic/extrinsic sensor calibration. You understand the math required to project a 3D Lidar point onto a 2D image pixel accurately.
- Early Fusion Expertise: Practical experience with architectures that fuse modalities at the feature level (e.g., BEVFusion, TransFuser, PointPainting) rather than just fusing final bounding boxes.
- SOTA Object Detection experience with modern transformer-based architectures (DETR, PETR, etc…) including similar temporal models (PETRv2, StreamPETR, …)
- Systems Fluency: You are an expert in Python, but you are also comfortable reading and writing systems code in C++ or Rust. You understand memory management and real-time constraints.
- Data Intuition: You understand that in robotics, better data alignment often beats a bigger model. You are willing to dig into the data infrastructure to ensure ground truth quality.