Perception engineer
Eklavya
Location: Onsite (Bangalore)
Role Summary
We are looking for a highly capable Perception Engineer to build robust perception systems for outdoor robotics. This role focuses on developing vision models, data pipelines, and inference systems for challenging real-world environments, with strong emphasis on accuracy, edge-case handling, maintainability, and deployment readiness. The ideal candidate should combine strong fundamentals with applied vision engineering expertise across image processing, deep learning, data-centric model development, failure analysis, and deployment optimization. We are looking for someone who can understand complex problems deeply, define clear evaluation standards, and build reliable perception pipelines for real-world operation.
Key Responsibilities
- Design and develop perception pipelines for real-world environments using computer vision, deep learning, and geometric reasoning.
- Build and improve models for scene understanding, recognition, segmentation, tracking, and spatial perception using modern deep learning architectures.
- Develop perception systems for monocular scene understanding and depth-related tasks where explicit depth sensing may not be available.
- Develop and evaluate online CNN models for human detection, tracking, and path prediction under occlusion for safe robotic operation.
- Translate use cases into measurable ML objectives, success metrics, and acceptance criteria such as mIoU, mAP, precision/recall, latency, FPS, and failure-case coverage.
- Own the data pipeline end to end: data collection planning, annotation workflow definition, dataset curation, pre-processing, augmentation, label quality checks, and post-processing.
- Work with annotation teams and tools such as CVAT to create high-quality datasets, including pixel-wise annotations, taxonomy definition, and edge-case labeling.
- Fine-tune and adapt modern foundation models such as SAM, DINOv2, and evaluate when they are useful versus custom task-specific architectures.
- Apply both classical and modern vision techniques where appropriate.
- Optimize models for embedded and edge platforms, with focus on compute efficiency, inference speed, memory constraints, and deployment practicality.
- Build clean, maintainable, modular code and scalable training/inference pipelines with good engineering practices and reproducibility.
- Collaborate with robotics and embedded teams to integrate perception outputs into navigation, planning, and control systems.
- Maintain documentation, experiment records, deployment specifications, and clear records of model limitations and known failure modes.
- Bachelor’s or Master’s degree in Computer Science, Electronics, Robotics, AI/ML, Applied Mathematics, or a related field.
- Strong hands-on experience in computer vision and deep learning, especially for real-world perception problems.
- Solid understanding of core computer vision fundamentals, including feature extraction, filtering, transformations, segmentation principles, and visual reasoning.
- Strong knowledge of modern CV tasks such as object detection, semantic/instance/panoptic segmentation, tracking, and scene perception.
- Experience building and training models using deep learning frameworks such as PyTorch or TensorFlow.
- Good understanding of architectures such as CNNs, Vision Transformers, encoder-decoder segmentation networks, and modern representation-learning approaches.
- Familiarity with concepts and methods such as SIFT, feature matching, geometric vision basics, DINO/DINOv2, SAM, or equivalent modern vision approaches.
- Experience in dataset creation, annotation workflows, pre-processing, post-processing, and quality validation for vision systems.
- Hands-on experience with annotation platforms such as CVAT.
- Strong programming skills in Python, with emphasis on clean code, modularity, and maintainability.
- Strong mathematical foundation in linear algebra, probability, optimization, and geometry, with comfort applying basic physics intuition where relevant.
- Ability to work across functions with robotics, embedded, and product teams in a fast-moving engineering environment.
- Experience building perception systems for outdoor, robotic, or embedded applications.
- Experience with fine-grained segmentation, visual tracking, spatial reasoning, or monocular scene understanding, representation learning, and transfer learning for low-data domains.
- Experience optimizing models for edge or embedded platforms.
- Exposure to ROS/robotics pipelines, deployment pipelines, or real-time inference systems.
- Familiarity with ONNX, TensorRT, quantization, pruning, profiling, and performance benchmarking.
- Experience with experiment tracking, reproducibility, and data/model versioning tools such as Weights & Biases, MLflow, DVC, or similar.
- Knowledge of reinforcement learning is a plus.
ARTPARK is India's leading deep-tech venture builder and incubator focused on robotics, connected autonomous systems, and AI. Leveraging our unique facilities and ecosystems, we strive to provide meaningful support to very early-stage startups building deep-tech products based in research. We are a non-profit organization created by Indian Institute of Science (IISc, Bengaluru) with support from the Department of Science & Technology (Government of India) and the Government of Karnataka.