New opportunity

Software Engineer, Perception

Intrinsic · Singapore

About this role

Intrinsic, an AI robotics group at Google, is seeking a Software Engineer for their Perception team to advance industrial robotics through AI, perception, and simulation. The role focuses on implementing, optimizing, and deploying robust perception capabilities to enable intelligent robot interaction in real-world industrial settings. Responsibilities include developing deep learning models for 3D computer vision tasks such as object detection, pose estimation, and 3D reconstruction, as well as optimizing algorithms for reliable grasping and manipulation. You will contribute to the design and documentation of core perception software and support production applications. Required qualifications include a 4-year degree in Computer Science, Electrical Engineering, Robotics, or a related field, and 1-2 years of experience in AI, machine learning, computer vision, or robotics with a focus on deep learning and transformer models. Proficiency in Python or C++ and at least one deep learning framework (PyTorch, TensorFlow, or JAX) is required, along with demonstrated experience in deploying computer vision solutions. Preferred qualifications include expertise in robotics, GPU kernel programming with CUDA, and experience in model optimization or quantization for deployment.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Deep Learning3D Computer VisionObject DetectionObject Pose Estimation3D ReconstructionGraspingTransformer modelsPyTorchTensorFlowJAXPythonC++CUDAModel OptimizationModel QuantizationRobotic Perception

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Deployment of 3D Computer Vision Models
The role requires translating deep learning research into production-ready software. You must demonstrate the ability to move beyond model training to implementing robust solutions for object detection, pose estimation, and 3D reconstruction in real-world industrial environments.
Optimization for Robotic Hardware
A core priority is ensuring perception algorithms perform reliably on robotic systems. This involves optimizing models for latency and accuracy, with a specific preference for candidates who can apply model quantization or GPU kernel programming via CUDA to meet real-time constraints.
Integration of Perception with Manipulation
You will be responsible for enabling intelligent robot interaction. This requires bridging the gap between raw perception data and actionable grasping or manipulation tasks, ensuring the software can handle the complexities of dynamic manufacturing settings.

Likely challenges to think through

Bridging Research and Industrial Reliability
Industrial settings are often less controlled than research environments. Prepare to discuss how you handle edge cases, sensor noise, and environmental variability when deploying deep learning models, as the primary challenge is maintaining high reliability in non-ideal conditions.
Balancing Model Complexity with Real-time Constraints
Using advanced transformer models often comes with high computational costs. You may face the challenge of needing to simplify or optimize these architectures to run efficiently on edge hardware, requiring a deep understanding of the trade-offs between model performance and inference speed.

Practice your answers

Technical practice

Deep Learning for 3D Vision

What to coverExplain how you would approach 3D object detection or pose estimation using deep learning. Discuss the trade-offs between different architectures, such as point-based versus voxel-based methods, and how they handle sensor noise or occlusion in industrial settings.

Transformer Models in Perception

What to coverDescribe your experience applying transformer-based models to computer vision tasks. Discuss how attention mechanisms improve performance over traditional CNNs for spatial feature extraction or temporal sequence modeling in robotics.

Model Optimization and Quantization

What to coverExplain the process of quantizing a deep learning model for deployment on edge hardware. Discuss the impact of precision reduction (e.g., FP32 to INT8) on model accuracy and inference latency, and how you mitigate potential performance degradation.

GPU Acceleration with CUDA

What to coverDescribe how you have utilized CUDA to optimize perception algorithms. Focus on your understanding of memory management, thread hierarchy, and how to identify bottlenecks in custom kernels for real-time robotic applications.

Robotic Grasping Challenges

What to coverDiscuss the technical challenges of implementing reliable grasping in dynamic environments. Address how you integrate perception data, such as point clouds or depth maps, into a control pipeline to ensure robust manipulation.

Framework Proficiency (PyTorch/TensorFlow/JAX)

What to coverCompare the strengths of your preferred deep learning framework regarding research flexibility versus production deployment. Suggestion: Be prepared to discuss how you manage data pipelines and model versioning within your chosen framework.

3D Reconstruction Techniques

What to coverExplain the methodologies you have used for 3D reconstruction, such as structure-from-motion or depth-from-stereo. Discuss how you handle calibration errors and environmental lighting variations to maintain reconstruction fidelity.

Software Engineering for Robotics

What to coverHow do you ensure the robustness and maintainability of perception software in a production environment? Discuss your approach to documentation, unit testing for computer vision modules, and handling edge cases in real-world sensor data.

Behavioural practice

Translating Research to Production

What to coverDescribe a situation where you had to adapt a research-level deep learning model for a production environment. Focus on the specific constraints you faced, the trade-offs you made, and the final result regarding performance and reliability.

Collaborating with Cross-functional Teams

What to coverProvide an example of working with roboticists or hardware engineers to integrate your perception software. Explain how you communicated technical requirements and resolved integration issues to meet project goals.

Handling Technical Ambiguity

What to coverDescribe a time when you were tasked with a perception problem with limited data or unclear requirements. Explain the steps you took to define the scope, gather necessary information, and deliver a functional solution.

Supporting Production Applications

What to coverShare an experience where you had to troubleshoot a perception failure in a deployed system. Detail your diagnostic process, how you identified the root cause, and the steps taken to prevent future occurrences.

Sharpen your knowledge

Topics to prepare

3D Computer Vision and Pose Estimation
Review the mathematical foundations of 6DoF pose estimation and 3D reconstruction. Exercise: Implement a pipeline that estimates the pose of a known object from a point cloud or RGB-D input. This is directly relevant to the role's requirement for enabling intelligent robot interaction in industrial settings.
Transformer Models in Vision
Study the application of Vision Transformers (ViT) and their variants for spatial reasoning. Exercise: Analyze how transformer-based architectures differ from traditional CNNs in object detection tasks. This addresses the explicit requirement for experience with transformer models in a perception context.
Robotic Grasping and Manipulation
Understand the intersection of perception and control for robotic grasping. Exercise: Research common approaches for mapping perception outputs to grasp candidates. This is essential for the role's focus on optimizing algorithms for reliable manipulation.

Skills to revise

Model Optimization and Quantization
Practice techniques to reduce model latency for real-time deployment. Exercise: Take a pre-trained PyTorch model and apply post-training quantization to observe the trade-off between inference speed and accuracy. This is a preferred qualification for deployment-focused roles.
CUDA Programming
Gain familiarity with writing custom GPU kernels to accelerate perception pipelines. Exercise: Write a simple CUDA kernel for a common image processing operation like a Gaussian blur or color space conversion. This is a highly valued skill for optimizing perception software at Intrinsic.
Deep Learning Framework Proficiency
Ensure fluency in PyTorch, TensorFlow, or JAX for end-to-end model development. Exercise: Build and train a custom object detection model from scratch using your preferred framework, focusing on clean code and modular design. This is a core requirement for the role.

Your preparation checklist

Prepare a 3D Computer Vision Project Walkthrough
Select a past project involving object detection, 6DoF pose estimation, or 3D reconstruction. Prepare to explain your model architecture, specifically focusing on your use of transformer-based models or deep learning frameworks like PyTorch, TensorFlow, or JAX, and how you handled data processing for 3D environments.
Demonstrate Model Deployment and Optimization Skills
Be ready to discuss your experience in taking a research-level model to a production environment. Prepare examples of how you optimized or quantized models to run efficiently on hardware, and if applicable, describe any experience you have with CUDA or GPU kernel programming to improve inference speed.
Review Robotic Perception and Grasping Fundamentals
Brush up on the intersection of computer vision and robotics. Prepare to discuss how perception data is translated into robotic actions, specifically regarding grasping and manipulation tasks in dynamic industrial settings.
Formulate Technical Questions for the Perception Team
Prepare thoughtful questions regarding Intrinsic's perception stack, such as how they handle sensor fusion in industrial environments, the specific challenges they face with real-time inference on robotic hardware, or how their simulation-to-reality pipeline integrates with their perception models.

Availability has not yet been verified. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in Singapore. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…