I’m Kyle Vedder

I am graduating with my CS PhD from Penn in May 2025, and I am on the job market for an industry Research Scientist position, either at a major AI lab or at a robotics startup, with the goal of working towards an embodied AI system that can perform at or above human level in a variety of domestic and industrial labor tasks.

Research Interests

I believe the shortest path to getting robust, generally capable robots in the real world is through the construction of systems whose performance scales with compute and data, without requiring human annotations.

In service of this, I am interested in designing and scaling fundamentally 3D vision systems that learn just from raw, multi-modal data. My contrarian bet is on the multi-modal and 3D aspects; a high quality, 3D aware representation with diverse data sources should enable more sample efficient and robust downstream policies. Most representations today are 2D for historical reasons (e.g. lots of RGB data, 2D convolutions won the hardware lottery), but I believe this ends up pushing a lot of 3D spacial understand out of the visual representation and into the downstream policy, making them more expensive to learn and less robust.

My current line of work is focused on tackling scene flow, a problem that requires systems to construct a robust understanding of the dynamics of the 3D world. For data availability reasons, it primarily focuses on the Autonomous Driving domain, but the same principles apply to other domains, e.g. indoor service robots.

Academic Background

I am a CS PhD candidate at Penn under Eric Eaton and Dinesh Jayaraman in the GRASP Lab. My current line of work is focused on Scene Flow with the general goal of building flexible, scalable systems that do not require human annotations.

Representative projects include:

• ZeroFlow: Scalable Scene Flow via Distillation
  • Distill an expensive optimization method into a feed-forward network and data scale it to state-of-the-art performance
• I Can’t Believe It’s Not Scene Flow!
  • Standard benchmarks were systematically broken, hiding the failure of all methods to describe small object motion; we proposed a new metric and a simple baseline that was state-of-the-art
• Argoverse 2 2024 Scene Flow Challenge
  • Hosted a challenge to push the field to close the gap between prior art and a qualitative notion of reasonable flow quality
• Scene Flow as a Partial Differential Equation
  • Fit a neural network-based scene flow volume to the entire observation and optimize it against multi-frame objectives and out pops amazing performance and emergent behaviors like point tracking.

For a narrative overview of how my PhD research fits together, see Overview of my PhD Research.

Industry / Engineering Background

I have done many industry internships:

• NVIDIA as a research intern under Zhiding Yu
• ArgoAI as a research intern under James Hays
• Amazon Lab126 as a perception intern on their home robot, Astro
• Google as a software engineering intern on adwords and ads quality
• Unidesk (a startup since aquired by Citrix)

I also have significant experience doing high-precision full stack robotics. In undergrad, I lead the greenfield development of AMRL’s Robocup Small Size League control stack and did research in multi-agent path planning.

More Information

• Email: kvedder (at) seas.upenn.edu
• Resume: /resume
• Publications: /publications
• GitHub: kylevedder
• Google Scholar: Kyle Vedder
• Twitter: KyleVedder

Updates