Senior Algorithm Engineer - Perception, 3D Reconstruction & Autonomy

Location: On-site, near-Raanana, with regular access to the robot and field work

About the role

At TreeX we’re building an autonomous agricultural robot that flies to a vine, understands the plant, decides where to cut, and makes the cut.

That means our algorithms don’t live only in notebooks or benchmarks. They run on real hardware, outdoors, under changing light, motion, occlusion, sensor noise, and all the other things the physical world throws at you.

We’re looking for a senior algorithm engineer who enjoys owning problems end to end: understanding the underlying math, building a robust implementation, validating it on real data, deploying it to the robot, and figuring out what happened when reality disagrees with the model.

This is a hands-on role for someone who likes going deep technically while still thinking about the full system.

What you’ll work on

You’ll be part of the team responsible for turning raw sensor data into reliable decisions about the plant.

Depending on the problem, that may include:

• 3D reconstruction from depth cameras - building and improving multi-camera point-cloud pipelines, world-frame aggregation, TSDF fusion, weighting, filtering, and drift detection.
• Scan alignment and pose - registration between scans, reconstruction and SLAM, along with confidence measures that are useful enough to drive downstream behavior.
• SLAM and state estimation - visual/inertial pose, sensor latency, EKF gating, transforms, calibration, and figuring out whether an error came from perception, pose, timing, or something else entirely.
• Semantic perception on the edge - training and deploying segmentation models on our own field data, exporting through ONNX/TensorRT, and running them on Jetson under real latency constraints.
• Datasets and training loops - annotation workflows, dataset versioning, benchmark sets, data curation, error analysis, and deciding what data is actually worth labeling next.
• Plant structure extraction - turning a reconstructed and labeled point cloud into trunks, canes, spurs, branches, relationships, and geometric features.
• Decision algorithms - scoring candidate cut points, defining validity constraints and grouping rules, encoding agronomic objectives, and building confidence measures that know when the robot should not act.
• Motion and manipulation at the interface - working closely with kinematics, planning, tracking, calibration, and collision constraints when the perception and motion worlds meet.

Our main stack includes modern C++17/20, Python, ROS 2, Jetson, TensorRT, CUDA, Eigen, PCL, OpenCV, Open3D, and OMPL. We also use Rust and TypeScript in other parts of the system.

What we’re looking for

Core experience

• Around 6+ years of experience developing algorithms for production systems.
• Significant ownership of at least one non-trivial perception, 3D geometry, estimation, or related algorithmic subsystem.
• Strong modern C++ skills, including performance profiling and optimization of real-time or near-real-time pipelines.
• Strong foundations in 3D geometry: coordinate frames, rigid transforms, quaternions, covariance, registration, projective geometry, and calibration.
• Practical machine-learning experience, including training, evaluating, and deploying models on real datasets.
• Experience working within edge or embedded constraints such as latency, memory, power, thermals, or model-export limitations.
• A solid approach to experimental validation: meaningful metrics, representative test sets, confidence in results, and awareness of leakage and variance.
• Clear technical communication in English, both written and spoken.

Especially useful experience

Any of the following would be a strong plus:

• ROS 2, including TF, QoS, executors, lifecycle, and debugging distributed robotic systems.
• CUDA or TensorRT optimization beyond basic model export.
• SLAM, factor graphs, sensor fusion, or state estimation.
• Manipulator kinematics or sampling-based motion planning.
• Experience in agriculture, medical devices, autonomous systems, or another domain where software decisions have physical consequences.
• Comfort working outside your primary language or subsystem when that’s where the real problem is.

You’ll probably enjoy this role if

• You like getting a messy real-world problem rather than a perfectly specified ticket.
• You want to understand why something failed instead of adding a patch around it.
• You care about both the algorithm and the software that makes it run reliably.
• You’re happy moving between Python experiments, C++ implementation, robot logs, point clouds, model outputs, and field testing.
• You like working with people who will challenge your reasoning and expect you to challenge theirs.
• You get satisfaction from seeing an algorithm you built change the behavior of a physical machine.