ML Developer for the Detector Pretraining Subgroup in Autonomous Transport

The task of the scene perception team for an autonomous vehicle is to transform data from various sensors into a world description sufficient for the operation of other components.

The detector pretraining subgroup researches methods for obtaining models with high generalizability. The model's output vector—a high-level scene representation—enables neighboring teams to both accelerate experiments and modify the vehicle's final motion trajectories.

What tasks await you

Experimentation with pretraining and alignment frameworks We want to experiment both with purely sensory pretrains, trained in an unsupervised setting, and with linking to text representations.

Compressing scene embedding to compact sizes Different consumers require different vector sizes, and it is necessary to obtain honest estimates of the trade-off between detail and compactness.

Preparing the training infrastructure The data is all trips collected by the fleet. For effective work on data from different modalities across different sensor sets, good infrastructure design is needed.

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

We expect that you

• Train neural network models, work with different approaches
• Have worked with PyTorch, TensorFlow, and frameworks built on top of them
• Strive to use and disseminate good, universal, scalable practices in your work
• Are familiar with the world of applied DL: know computer vision theory, achieve practical results
• Have worked with large models (LLM, VLM, Foundation Models) in other domains (will be an advantage)
• Read papers from ML conferences, follow developments in the field