Back
Learning Dexterity: Dactyl Robot Hand Manipulation via Sim-to-Real Transfer
webCredibility Rating
4/5
High(4)High quality. Established institution or organization with editorial oversight and accountability.
Rating inherited from publication venue: OpenAI
A 2018 OpenAI capabilities milestone relevant to AI safety discussions around sim-to-real generalization, robustness, and the gap between simulated training environments and real-world deployment conditions.
Metadata
Importance: 52/100blog postprimary source
Summary
OpenAI presents Dactyl, a system that trains a Shadow Dexterous Hand robot entirely in simulation using reinforcement learning, then transfers the learned policy to a physical robot without fine-tuning. The system achieves unprecedented dexterous object manipulation by solving challenges including high-dimensional control, noisy observations, and sim-to-real transfer gaps. This demonstrates that physically-accurate world modeling is not required for real-world task performance.
Key Points
- •Dactyl trains entirely in simulation using the same RL algorithm as OpenAI Five, then deploys to a real robot hand with 24 degrees of freedom without fine-tuning.
- •Sim-to-real transfer is achieved through domain randomization, adapting to real-world physics without precise physical modeling of contact or friction.
- •The system handles noisy, partial observations from fingertip sensors and RGB cameras, inferring unobservable quantities like friction and slippage.
- •Generalizes across multiple object geometries (blocks, prisms), showing the approach is not limited to task-specific strategies.
- •Represents a milestone in dexterous robotic manipulation, an area where traditional robotics approaches had seen limited progress.
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| Goal Misgeneralization Probability Model | Analysis | 61.0 |
Cached Content Preview
HTTP 200Fetched Apr 9, 202615 KB
Learning dexterity
Mar
APR
May
10
2023
2024
2025
success
fail
About this capture
COLLECTED BY
Collection: Save Page Now
TIMESTAMPS
The Wayback Machine - http://web.archive.org/web/20240410030903/https://openai.com/research/learning-dexterity
Close
Search Submit
Skip to main content
Site Navigation
Research
Overview
Index
GPT-4
DALL·E 3
Sora
API
Overview
Pricing
Docs
ChatGPT
Overview
Team
Enterprise
Pricing
Try ChatGPT
Safety
Company
About
Blog
Careers
Residency
Charter
Security
Customer stories
Search
Navigation quick links
Log in
Try ChatGPT
Menu
Mobile Navigation
Close
Site Navigation
Research
Overview
Index
GPT-4
DALL·E 3
Sora
API
Overview
Pricing
Docs
ChatGPT
Overview
Team
Enterprise
Pricing
Try ChatGPT
Safety
Company
About
Blog
Careers
Residency
Charter
Security
Customer stories
Quick Links
Log in
Try ChatGPT
Search
Submit
Illustration: Ben Barry & Eric Haines
Research
Learning dexterity
We’ve trained a human-like robot hand to manipulate physical objects with unprecedented dexterity.
July 30, 2018
More resources
Read paper
Robotics, Domain randomization, Meta-learning, Sim-to-real, Transfer learning, Dactyl, Milestone
Learning dexterity3:29
Our system, called Dactyl, is trained entirely in simulation and transfers its knowledge to reality, adapting to real-world physics using techniques we’ve been working on for the past year. Dactyl learns from scratch using the same general-purpose reinforcement learning algorithm and code as OpenAI Five. Our results show that it’s possible to train agents in simulation and have them solve real-world tasks, without physically-accurate modeling of the world.
Examples of dexterous manipulation behaviors autonomously learned by Dactyl.
The task
Dactyl is a system for manipulating objects using a Shadow Dexterous Hand. We place an object such as a block or a prism in the palm of the hand and ask Dactyl to reposition it into a different orientation; for example, rotating the block to put a new face on top. The network observes only the coordinates of the fingertips and the images from three regular RGB cameras.
Although the first humanoid hands were developed decades ago, using them to manipulate objects effectively has been a long-standing challenge in robotic control. Unlike other problems such as locomotion, progress on dextrous manipulation using traditional robotics approaches has been slow, and current techniques remain limited in their ability to manipulate objects in the real world.
Reorienting an object in the hand requires the following problems to be solved:
Working in the real world. Reinforcement learning has shown many successes in simulations and video games, but has had comparatively limited results in the real world. We test Da
... (truncated, 15 KB total)Resource ID:
6ff01553c3d5a60f | Stable ID: sid_T9LronxPMp