Robotics AI Engineer | Defence in Noida, भारत

Role Name -  Robotics AI Engineer

Experience- 2-3 years

Location -  Sector 132, Noida 

About ZOID 

ZOID is a deep-tech R&D startup working on advanced technologies for the defence and aerospace sector. We build customised solutions for the Indian Military and UAV manufacturers, with our strength being hands-on expertise in AI, electronics, UAVs and robotics.

Current projects include: 

• AI-based FOD detection software for naval air stations 

• Reusable Off-board Missile Decoy for anti-ship missile defence 

• GNSS-Denied Navigation Suite 

• Aerial Intelligent Mapping Suite 

• Vision-based Kinetic Strike Solution 

• Swarming Solutions 

Role Brief: 
We are looking for an Robotics AI Engineer to handle AI data operations through field data collection, image annotation, dataset QC, and test-set analysis for UAV/defence applications. The role involves working with field teams during runway/track tests and maintaining accurate, traceable datasets and test logs. 

Key Responsibilities: 

• Own the optics testing end to end (FOV, focal length and depth-of-field calculations) for detecting 3 x 3 x 3 mm objects. 

• Plan and conduct optics testing including camera & lens settings (focus, motion-blur budget, exposure, gain, FPS, binning, etc).

• Build and manage the dataset, annotation and QA/QC rules, data versioning, and backup (Raid Configuration). 

• Analyse the type of datasets which should be collected for any specific task.

• Develop the detection & classification model for millimetre-scale objects at high resolution. 

• Optimise models for real-time inference on the GPU servers such that multiple cameras can simultaneously operate on a single GPU without any accuracy loss. 

• Implement object localisation and geo-referencing of detections using RTK/GNSS and frame timestamps and validate CEP in the field.

• Maintain a model registry with regression tests so no release degrades detection performance; run the false-positive review loop with operations staff.

• Analyse why false positives are occurring and how to eliminate them?

• Recommend computation hardware (GPU, memory, storage, camera network cards) based on model and data-rate requirements.

• Apply AI methods to 3D photogrammetry data. Run segmentation, object detection and change detection, and extend the 3D processing pipeline where required. (Good to have) 

Must Have: 

• Fluency in Python. 

• Strong computer vision fundamentals 

• Knowledge of PyTorch & implementation of object detection at high resolution and small object scale (YOLO/DETR family or equivalent).

• Dataset engineering: annotation tooling (CVAT/Label Studio), QC methods, data versioning (DVC), experiment tracking (W&B/MLflow).

• TensorRT, CUDA, ONNX, INT8 quantisation, multi-GPU inference.

• Optics knowledge: FOV, IFOV, GSD, focal length, MTF, exposure & experience with machine-vision cameras and NIR imaging (Good to have) 

• Object localisation and geo-referencing: pixel-to-ground mapping, CEP estimation, working with RTK/GNSS and time-synchronised frames. (Good to have) 

• 3D data handling: point clouds and meshes (Open3D, PCL, PDAL or equivalent), knowledge of 3D deep learning models (PointNet/PointNet++, KPConv or similar) for segmentation and detection on 3D models. (Good to have) 

• Linux, Git and CI; comfortable deploying and debugging on field hardware.

Good to Have: 

• Experience with multi-camera systems: synchronisation, overlapping fields of view, and stitching or fusing outputs from a camera array.

• Familiarity with GigE-Vision/GenICam camera SDKs and PoE camera networks, enough to debug frame drops and timing with the software team.

• Active learning or semi-automatic labelling to cut annotation effort on large field datasets.

• Exposure to anomaly detection or unsupervised methods for finding objects that were never in the training set.

• LiDAR or depth-sensor data handling, useful for cross-checking optical detections and for AIMS 3D work.

• Understanding of EO/IR sensor behaviour beyond NIR (SWIR/MWIR), for future sensor upgrades.

• Basic Docker and model-serving experience for reproducible field deployments.