2018

July

1 events

Editors' summary

A hand trained only in simulation turns a block in the real world

On 30 July, OpenAI announced Dactyl, a robot hand system trained entirely in simulation, without data from the physical robot, that went on to reorient a block in the real world.

Randomising the conditions across a vast number of trials absorbed the gap between simulation and reality — an attempt at getting there without going out to collect the data.

This block is written by the editors. It is kept separate from the sourced record below.

Record1 events
  1. 30
    ResearchOpenAI

    Dactyl robot hand system announced

    OpenAI announced Dactyl, a system that manipulates objects with a human-like robot hand (the Shadow Dexterous Hand). Trained entirely in simulation, the learned policy transferred to the physical robot without fine-tuning. Using the same general-purpose reinforcement learning algorithm and code as OpenAI Five, Dactyl reoriented a block in the palm of the hand, achieving up to 50 consecutive rotations, demonstrating that agents trained in simulation can solve real-world tasks.