What Matters When Cotraining Robot Manipulation Policies on Everyday Human Videos?

TL;DR

Study co-trains robot manipulation policies using everyday human videos, achieving a 29.7% success rate improvement.

cs.RO 🟡 Intermediate 2026-06-05 41 views
Richard Li Aditya Prakash Andrew Wen Saurabh Gupta Yilun Du Pulkit Agrawal
robotics human videos cotraining hand pose estimation dataset

Key Findings

Methodology

The study uses 532 human videos and 3,000 robot demonstrations, obtaining high-quality 3D hand pose labels through multi-view triangulation. It addresses the motion gap between humans and robots by specializing vision and policy networks.

Key Results

  • Achieved a 29.7% success rate improvement in low-robot-data scenarios.
  • Everyday video data showed better robot transfer than lab data.
  • Hand pose quality significantly affects transfer, with triangulation outperforming monocular estimation.

Significance

This research offers a new approach to training robot manipulation policies using everyday internet videos, addressing data scarcity and highlighting the impact of hand pose quality and motion gap on transfer effectiveness.

Technical Contribution

Introduces a novel cotraining strategy that significantly improves the transfer of robot manipulation policies by specializing vision and policy networks.

Novelty

First systematic study on using everyday video data for robot manipulation policy training, focusing on hand pose quality and motion gap impacts.

Limitations

  • Hand pose estimation accuracy remains a bottleneck affecting transfer effectiveness.
  • Motion transfer is less evident in complex tasks.
  • Larger datasets are needed for validation.

Future Work

Future research could focus on improving hand pose estimation quality and exploring new representational alignment and action retargeting techniques.

AI Executive Summary

Training robot manipulation policies faces data scarcity challenges. Existing methods rely on curated demonstrations, which are difficult to scale. Researchers propose a new method using everyday internet videos, obtaining high-quality 3D hand pose labels through multi-view triangulation and addressing the motion gap between humans and robots by specializing vision and policy networks.

Experimental results show a 29.7% success rate improvement in low-robot-data scenarios. Everyday video data demonstrated better robot transfer than lab data, highlighting the significant impact of hand pose quality on transfer effectiveness. The study also found that simple image-space scale alignment can significantly improve transfer.

Despite significant progress, hand pose estimation accuracy remains a bottleneck, especially in complex tasks. Future research could focus on improving hand pose estimation quality and exploring new representational alignment and action retargeting techniques to further enhance the transfer of robot manipulation policies.

Deep Analysis

Background

Training robot manipulation policies typically relies on curated demonstrations, requiring expensive equipment and significant time to collect. Internet videos offer a potentially scalable data source, but effectively utilizing this data remains a challenge.

Core Problem

The core problem is effectively extracting information from everyday human videos to train robot manipulation policies. Existing methods face bottlenecks in motion gap and hand pose estimation quality.

Innovation

The study proposes a novel cotraining strategy that addresses the motion gap between humans and robots by specializing vision and policy networks.

Methodology

  • �� Utilized 532 human videos and 3,000 robot demonstrations.
  • �� Obtained high-quality 3D hand pose labels through multi-view triangulation.
  • �� Specialized vision and policy networks for effective transfer.

Experiments

The experiments used 532 human videos and 3,000 robot demonstrations to evaluate transfer effectiveness across six real-world manipulation tasks. The study compared different hand pose estimation methods to analyze the impact of hand pose quality.

Results

Experimental results showed a 29.7% success rate improvement in low-robot-data scenarios. Everyday video data demonstrated better robot transfer than lab data.

Applications

The method can be used for training robot manipulation policies, especially in data-scarce scenarios. It provides a new approach for utilizing internet videos in robot training.

Limitations & Outlook

Hand pose estimation accuracy remains a bottleneck affecting transfer effectiveness. Motion transfer is less evident in complex tasks.

Plain Language Accessible to non-experts

Imagine you're in a kitchen learning to cook by watching others chop, stir, etc. Robots can learn operations by watching human videos too, but they need very precise hand data to mimic human actions. Researchers used multi-view triangulation to obtain high-quality hand data and specialized training strategies to help robots better learn these actions.

ELI14 Explained like you're 14

Imagine playing a game where you have to mimic a character's moves. Robots have a similar task—they learn to operate by watching human videos. Researchers found a smart way to help robots understand and mimic these moves better, just like you practice in a game until you master it!

Glossary

Cotraining

A machine learning method that uses multiple data sources to improve model performance.

Used for training robot manipulation policies.

Hand Pose Estimation

A technique for identifying and tracking hand positions and postures.

Used to obtain high-quality 3D hand pose labels.

Triangulation

A technique for calculating object positions using multiple viewpoints.

Used to obtain high-quality 3D hand pose labels.

Vision and Policy Networks

Neural networks used for processing visual information and formulating operation strategies.

Used for specialized training.

Motion Gap

The difference between human and robot actions.

A key factor affecting transfer effectiveness.

Open Questions Unanswered questions from this research

  • 1 How to improve hand pose estimation accuracy to enhance transfer effectiveness?
  • 2 How to achieve better motion transfer in complex tasks?

Applications

Immediate Applications

Robot Manipulation Training

Utilizing everyday human videos to train robot manipulation policies, especially in data-scarce scenarios.

Long-term Vision

Intelligent Robots

Developing intelligent robots capable of learning complex tasks from internet videos.

Abstract

Human video datasets used for cotraining robot manipulation policies largely consist of curated demonstrations where motions are orchestrated to resemble robot behavior and 3D hand poses are captured with specialized hardware. A more plentiful source of data is everyday Internet video, but it is an open question what factors enable transfer from such videos to robots. We investigate this using a new dataset of 532 human videos with 28 hours of high-quality triangulated hand labels and natural motions. We find that hand pose quality affects transfer, but even with accurate hands, the inherent motion gap hinders transfer unless the vision and policy networks specialize to each embodiment. Our cotraining recipe yields consistent improvements, with an absolute success rate gain of $29.7\%$ in the low-robot-data regime across six manipulation tasks.

cs.RO cs.AI cs.CV cs.LG