Shuran Song

Stanford Professor, Diffusion Policy Creator

Shuran Song

Home > People > Shuran Song

Profile

FieldDetails
Current PositionAssistant Professor, Stanford University
PreviousAssistant Professor, Columbia University (2019-2024)
PhDPrinceton University
LabColumbia Artificial Intelligence and Robotics Lab (CAIR)

Key Contributions

  • Diffusion Policy: Applied diffusion to robot action generation, new paradigm for VLA action generation
  • 3D Perception: 3D perception research for robotics
  • UMI (Universal Manipulation Interface): Universal data collection interface
  • Columbia to Stanford Move: Strengthening robot learning research

Research Timeline

PhD & Postdoc (2013-2019)

Princeton -> Postdoc

YearWorkImpact
20153D ShapeNetsEarly 3D deep learning research
2017Semantic Scene Completion3D scene understanding
2018PhD Graduation

Columbia University (2019-2024)

CAIR Lab Founding

YearWorkImpact
2019Joined Columbia as ProfessorFounded CAIR Lab
2021Transporter Networks collaborationObject rearrangement
2023Diffusion PolicyPioneering robot diffusion research
2024UMIUniversal manipulation interface

Stanford University (2024-present)

Moved to Stanford

YearWorkImpact
2024Moved to Stanford as Professor
2024-Continuing robot learning research

Major Publications

Diffusion for Robotics

  • Diffusion Policy (RSS 2023, IJRR 2024) - Pioneering robot diffusion research
  • 3D Diffusion Policy (2024)

3D Perception

  • 3D ShapeNets (CVPR 2015)
  • Semantic Scene Completion (CVPR 2017)
  • ScanNet (CVPR 2017)

Robot Manipulation

  • UMI (Universal Manipulation Interface, 2024)
  • Transporter Networks related research

Key Ideas

Diffusion Policy (2023)

Core: Model robot action generation as a denoising diffusion process

Noise -> ... -> Action sequence
(gradual denoising)

Advantages:
- Handling multimodal action distributions
- High training stability
- Suitable for high-dimensional action spaces

Impact:

  • Influenced pi0 (flow matching), Octo (diffusion decoder), and others
  • LeRobot default supported model
  • New paradigm for robot action generation

UMI (Universal Manipulation Interface, 2024)

Core: Universal robot data collection interface

Features:
- Applicable to various robot platforms
- Low-cost data collection
- Standardized interface

Philosophy & Direction

Research Philosophy

“3D world understanding and robot manipulation are closely connected”

Research Direction Evolution

  1. 2013-2018: 3D deep learning, scene understanding
  2. 2019-2022: 3D perception for robotics
  3. 2023-present: Diffusion for robot learning, manipulation interfaces

Key Collaborations

  • Cheng Chi: Diffusion Policy lead author, UMI collaboration
  • Toyota Research Institute: Diffusion Policy collaboration
  • MIT: Diffusion Policy collaboration

Awards & Recognition

  • NSF CAREER Award
  • Amazon Research Award
  • Google Research Scholar

See Also