Master Thesis: Monocular XR Object Localization with World Models

Lightweight XR and AI glasses may need to operate without dedicated depth sensors due to constraints on cost, energy use and size. This thesis at Ericsson explores whether a pretrained world model, such as NVIDIA Cosmos, can exploit semantic and scale cues to estimate the distance and motion of faraway objects in monocular XR video. Its performance will be compared with headset depth sensing and standard monocular depth-estimation methods.

The assignment

• Develop an RGB video capture pipeline for an XR headset, such as Quest 3, with inference running on a server.
• Train a lightweight probe to predict object distance and motion while leaving the pretrained world model backbone unchanged.
• Test the approach in controlled scenes featuring walking people and moving objects, comparing predictions with measured ground truth and depth-estimation baselines.
• Measure localization accuracy, computational demands, latency and uplink requirements.
• Deliver a functional prototype together with a documented dataset and measurement setup.
• Report quantitative accuracy and latency findings and provide recommendations for camera-only XR perception and network-assisted spatial intelligence.

Who we are looking for

• Master’s student with a strong academic focus on XR and spatial understanding.
• Good knowledge of Unity and MR/VR development.
• Experience developing applications for Meta Quest headsets is an advantage.
• Fluent in spoken and written English.
• Analytical, curious and receptive to new ideas.
• Aligned with Ericsson’s values of professionalism, perseverance, respect and integrity.

What we offer

• An opportunity to explore challenging technical problems and develop novel solutions within a team of innovators.
• A collaborative environment that encourages creativity and experimentation.

Practical details

Location: Stockholm. Reference number: 793531. Start date, duration and application deadline are not specified.