# Visual Imitation Enables Contextual Humanoid Control

**Authors:** Arthur Allshire, Hongsuk Choi, Junyi Zhang, David McAllister, Anthony Zhang, Chung Min Kim, Trevor Darrell, Pieter Abbeel, Jitendra Malik, Angjoo Kanazawa  
**arXiv:** [2505.03729](https://arxiv.org/abs/2505.03729)  
**Date:** 2025-05-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2505.03729v1/

## Key points

- Authors: Arthur Allshire, Hongsuk Choi, Junyi Zhang, David McAllister, Anthony Zhang, Chung Min Kim, Trevor Darrell, Pieter Abbeel, Jitendra Malik, Angjoo Kanazawa
- Paper Title: "Visual Imitation Enables Contextual Humanoid Control"
- Approach: VIDEOMIMIC
- Utilizes everyday videos to capture human motion and transfer knowledge to humanoid robots
- Real-to-sim-to-real pipeline for skill transfer
- Generates whole-body control policies for tasks like climbing stairs and sitting on chairs
- Demonstrates robust and repeatable performance on real humanoid robots in dynamic movements
- Scalable pathway for teaching humanoids to operate effectively in diverse environments
- Bridges gap between visual imitation learning and contextual understanding
- Paves way for adaptive robotic systems capable of complex tasks with ease

## Summary

In their paper titled "Visual Imitation Enables Contextual Humanoid Control," authors Arthur Allshire, Hongsuk Choi, Junyi Zhang, David McAllister, Anthony Zhang, Chung Min Kim, Trevor Darrell, Pieter Abbeel, Jitendra Malik, and Angjoo Kanazawa explore the challenge of teaching humanoids to perform complex tasks by leveraging contextual information from the surrounding environment. The authors propose a novel approach called VIDEOMIMIC that utilizes everyday videos to capture human motion and transfer this knowledge to humanoid robots through a real-to-sim-to-real pipeline. Through this method of reconstructing both humans and their environmental context, VIDEOMIMIC generates whole-body control policies that enable humanoid robots to autonomously replicate skills such as climbing stairs and sitting on chairs. The results of their pipeline demonstrate robust and repeatable performance on real humanoid robots in various dynamic movements. The authors emphasize that VIDEOMIMIC offers a scalable pathway for teaching humanoids to operate effectively in diverse real-world environments by bridging the gap between visual imitation learning and contextual understanding. This innovative approach paves the way for more adaptive and versatile robotic systems capable of navigating complex tasks with ease. Their research not only advances the field of robotics but also highlights the potential for integrating AI technologies into everyday scenarios to enhance human-robot interactions.