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

**Authors:** Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, Shuran Song  
**arXiv:** [2402.10329](https://arxiv.org/abs/2402.10329)  
**Date:** 2024-02-15  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2402.10329v1/

## Key points

- UMI is a data collection and policy learning framework for transferring skills from human demonstrations to robot policies in real-world settings
- UMI enables portable, cost-effective, and information-rich solutions for complex bimanual and dynamic manipulation tasks
- Key feature: deployable policy learning through a meticulously crafted policy interface with inference-time latency matching and relative-trajectory action representation
- Learned policies are hardware-agnostic and can be deployed across various robot platforms
- Enables new capabilities in robot manipulation, such as dynamic bimanual actions with long horizons by changing training data for specific tasks
- Versatility demonstrated through successful generalization to novel environments and objects in real-world experiments
- Hardware and software system of UMI is open-source at https://umi-gripper.github.io

## Summary

In their paper titled "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots," authors Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song introduce the groundbreaking . UMI is a data collection and policy learning framework that enables direct transfer of skills from human demonstrations in real-world settings to deployable robot policies. By utilizing hand-held grippers and carefully designed interfaces, UMI allows for portable, cost-effective, and information-rich  for complex bimanual and dynamic manipulation tasks. One of the key features of UMI is its ability to facilitate deployable policy learning through a meticulously crafted policy interface with inference-time latency matching and a relative-trajectory action representation. This design ensures that the learned policies are hardware-agnostic and can be deployed across various robot platforms. As a result, UMI unlocks new capabilities in robot manipulation by enabling  such as dynamic bimanual actions with long horizons. Remarkably, these behaviors can be achieved by simply changing the training data for each specific task. The authors demonstrate the versatility and efficacy of UMI through comprehensive real-world experiments where policies learned via UMI successfully generalize to novel environments and objects when trained on diverse human demonstrations. Furthermore, they have made the hardware and software system of UMI open-source at https://umi-gripper.github.io. Overall, by bridging the gap between human demonstrations in natural settings and deployable robot policies without requiring in-the-wild robots.