Cheng Chi

Diffusion Policy Lead Author, UMI Developer

Cheng Chi

Home > People > Cheng Chi

Profile

FieldDetails
Current PositionStanford University PhD
PreviousColumbia University
AdvisorShuran Song

Key Contributions

  • Diffusion Policy Lead Author: Applied diffusion to robot action generation
  • UMI (Universal Manipulation Interface): Universal manipulation data collection interface
  • 3D Diffusion Policy: Combined 3D representation with diffusion

Research Timeline

Columbia to Stanford PhD

Advised by Shuran Song

YearWorkImpact
2022Research BeginsColumbia CAIR Lab
2023Diffusion PolicyPioneering robot diffusion research
2024UMIUniversal manipulation interface
20243D Diffusion Policy3D + diffusion
2024Moved to StanfordTogether with Shuran Song

Major Publications

Diffusion Policy (RSS 2023, IJRR 2024)

“Diffusion Policy: Visuomotor Policy Learning via Action Diffusion”

Key contributions:

  • First application of diffusion to robot action generation
  • Handling multimodal action distributions
  • Average 46.9% performance improvement across 4 benchmarks

UMI (2024)

“Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots”

Key contributions:

  • Data collection with human hands, without robots
  • Transfer to various robot platforms
  • In-the-wild data collection

3D Diffusion Policy (2024)

  • 3D point cloud input
  • Diffusion policy + 3D representation

Key Ideas

Diffusion Policy (2023)

Core: Generate robot actions through denoising diffusion

p(a|o) = integral p(aK) product p(ak-1|ak, o) dak:K

Process:
1. Start from pure noise
2. Gradual denoising conditioned on observation (o)
3. Generate final action sequence

vs ACT:

AspectDiffusion PolicyACT
Generation MethodIterative denoisingSingle forward pass
MultimodalityNatural handlingStyle variable (z)
Training StabilityVery highHigh

UMI (2024)

Core: Collect data with human hands without robots, then transfer to robots

[Human hand demonstration] -> [UMI Interface] -> [Robot policy learning]

Advantages:
- Data collection without robot hardware
- Collection possible in real environments
- Transfer to various robots

Impact

Impact of Diffusion Policy

  • pi0: Adopted flow matching (a variant of diffusion)
  • Octo: Uses diffusion decoder
  • SmolVLA: Uses flow matching
  • LeRobot: Default supported model

Impact of UMI

  • Significantly reduced data collection costs
  • Influenced similar approaches like Sunday Robotics

Philosophy

Research Philosophy

“The key is combining good representations with good generative models”

Research Direction

  1. 2022-2023: Diffusion for robot learning
  2. 2024-present: Data collection interfaces, 3D representations

See Also