Research Sciences INTERN

Location: Switzerland, Zürich, Zürich Department: Research Sciences

Overview

The Spatial AI Lab is part of the Applied Sciences Group, a Microsoft research and development organization dedicated to creating next-generation human–computer interaction technologies, leveraging the latest AI developments and exploring new hardware capabilities and device form factors. Our scientists and engineers bring deep expertise in computer vision and multimodal AI, with a particular focus on spatial and embodied AI.

About the internship

We are seeking Research Interns in Computer Vision, Machine Learning and Robotics, broadly defined, for our Spatial AI Lab in Zurich. As an intern, you will collaborate with one or more mentors and use the cutting-edge hardware and software we are creating. During the 12-week internship, you will work on a stimulating, open-ended research project that seeks to advance the state of the art. An internship at Microsoft will not only enhance your career but also let you take part in exciting research breakthroughs. Interns work, learn, and network with some of the world’s top researchers and engineers.

This call covers four research directions, described below. The exact project will be shaped jointly by you and your mentors, based on research opportunities and your background. The start date of the internship could be the 1st of each month, no later than 1st of April 2027.

Microsoft’s mission is to empower every person and every organization on the planet to achieve more. As employees we come together with a growth mindset, innovate to empower others, and collaborate to realize our shared goals. Each day we build on our values of respect, integrity, and accountability to create a culture of inclusion where everyone can thrive at work and beyond.

Direction A – Multimodal Embeddings

Text–image embedding models such as CLIP and SigLIP are foundational building blocks for search, retrieval, and multimodal AI. In this direction, you will study how these models organize information in their embedding spaces and develop new methods that make embeddings more expressive, flexible, efficient, and interpretable, with applications such as semantic search.

Example research topics

Geometry and expressivity of multimodal embedding spaces (e.g., cone effect, modality gap)

Updating embedding models while preserving compatibility with existing embeddings

Multi-vector, hierarchical, and composable embeddings

Explainable and steerable embeddings

Generative search, efficient retrieval, and quantization for inference

Benchmarking retrieval beyond web-scale data, such as private collections

What we look for

Research experience in multimodal representation learning, text–vision embedding models, contrastive learning, neural information retrieval, or embedding adaptation and compression (e.g., knowledge distillation, quantization).

Direction B – Image Understanding and Editing

Advances in visual intelligence are creating new opportunities for models that interpret rich image content and reason about what makes an image meaningful. In this direction, you will investigate models that build a context-aware understanding of images (scene composition, geometry, appearance, salient regions, objects, and their relationships) and explore how this understanding can power downstream applications.

Richer visual representations and combining complementary visual signals

Context-aware scene and image understanding, and its use for image editing

Generalization, robustness, and reliable evaluation of visual understanding

Efficient architectures, model compression, or on-device deployment (optional extension)

Research experience in image understanding, visual representation learning, computational photography, multimodal learning, or generative image editing, with a strong background in modern discriminative, generative, or vision–language models. Experience with efficient neural networks is a plus.

Direction C – Robot Learning for Manipulation and Real-to-Sim-to-Real

Real-world observations and demonstrations are a rich source of experience for robot learning. In this direction, you will work on manipulation with stationary and mobile bimanual robots, exploring how physical experience, such as videos and demonstrations, can inform simulation, and how simulated experience can improve real-world policies.

Real-to-sim-to-real learning and policy robustness

Video–action models, world–action models, and vision–language–action (VLA) policies

Learning from human demonstrations, including 3D reconstruction and tracking of hand–object interactions

Physics-based and differentiable simulation for manipulation

Research experience in robot learning (imitation or reinforcement learning, sim-to-real transfer), bimanual or mobile manipulation, 3D reconstruction and pose estimation, or physics-based simulation; familiarity with tools like PyTorch, ROS, and MuJoCo; and experience deploying policies on real robots.

Note: Robot-learning projects in this direction involve hands-on work with physical robots and require presence at our Zurich lab for the full duration of the internship.

Direction D – Image Generation and Efficient Visual Generative Models

In this direction, you will explore compact and efficient visual generative models across image and video generation, and action-conditioned visual prediction. You will choose a focus based on your interests and background.

Video generation: extending image generation to coherent sequences conditioned on text or images

World modeling: lightweight models that predict future visual observations from context and actions

Architecture distillation: transferring the capabilities of large image generators into different, smaller architectures, including latent-to-pixel-space distillation

Research experience in generative model