Skip to content

Universal Manipulation Interface

Transfer human demonstrations into deployable robot policies

Visit official site

What Universal Manipulation Interface is

The framework is described as a data collection and policy learning system that moves skills from in-the-wild human demonstrations into deployable robot policies. It combines a handheld gripper with a designed policy interface, including inference-time latency matching and relative trajectory action representation.

What you can do with it

  • Collect portable data within two minutes of startup
  • Use camera-centric action representation without calibration
  • Collect each demonstration in about 30 seconds
  • Deploy zero-shot generalised policies in new environments

Application scenarios

  • Collect everyday actions such as ball tossing, folding clothes and dish washing
  • Deploy trained policies on different robot platforms without calibration
  • Use a CLIP-pretrained ViT as the visual encoder so policies react to changes

Focus areas

  • Robot skill learning
  • Handheld devices with external sensors
  • Human-robot interaction interface design

Pros & cons

✓ What we like

  • Each demonstration takes about 30 seconds to collect
  • Learned policies are hardware agnostic across robot platforms

! What to watch out for

  • Data collection depends on a handheld gripper with external sensors
  • Policy transfer is tied to the training demonstrations collected for each task

FAQ

What does UMI transfer?

The material states skills from in-the-wild human demonstrations are transferred into deployable robot policies.

Is calibration required?

The feature list states the camera-centric action representation needs no calibration.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review