Delivery Robot Perception Team Lead

Delivery robots are a step into the future of logistics and autonomous technology. We are building a system in which thousands of robots will be able to move independently around the city, interact with people and infrastructure, and deliver orders quickly and safely.

We have a short cycle: from idea to concept, from test site to production. There is a lot of data and enough GPU power for experiments. We are looking for a person who will be responsible for one of the perception directions.

You can read about our robot's design in the overview article on Habr.

What tasks await you

Improving recognition algorithms You will develop and improve neural network computer vision algorithms in the fields of 2D and 3D recognition and tracking. You will propose hypotheses, conduct experiments with accumulated data, and analyze results using existing and new quality metrics. Your goal is to create and implement architectures that accurately and compactly describe the surrounding world of the delivery robot.

Developing a Data Engine The larger and more diverse the data, the better the models become. You will develop an efficient Data Engine that automates the collection, selection, and labeling of data, reducing costs and accelerating the improvement cycle. You will work on tools and algorithms that will increase the quality and speed of labeling (including with the involvement of external annotators) and ensure maximum return from each new data example.

Experimental Technologies You will research new approaches to robot perception and world understanding, study foundation models and end-to-end approaches, and integrate new architectures into the robot's perception system. You will actively propose and test bold ideas that can radically increase the efficiency and reliability of our robots.

Model Optimization All our models run on the robot's onboard hardware in real time and with limited resources. Therefore, new solutions must be accurate and performant for execution on the robot. You will solve optimization tasks at various levels: from architectural decisions and efficient code writing to developing custom operations for TensorRT and other inference acceleration tools.

More about ML at Yandex — in the Yandex for ML channel

We expect that you

• Have managed a team
• Are perfectly oriented in the world of applied computer vision: you know the theory and methods of working with CV tasks, have successfully applied them, and achieved practical results
• Have integrated CV and ML solutions into products
• Have worked with PyTorch, understand the specifics of parallel computing, write efficient GPU code
• Program in Python or C++

It will be a plus

• Have solved 2D or 3D object detection and tracking tasks
• Have worked with sensors other than cameras: LiDARs, radars
• Are familiar with TensorRT, neural network acceleration, quantization, have written custom operations
• Have participated in artificial intelligence competitions or sports programming