Skip to content

DigiRL

Train device agents with online reinforcement learning

Visit official site

What DigiRL is

The algorithm is described as an online reinforcement learning method that trains agents to control devices on real-world Android tasks.

What you can do with it

  • Make use of existing offline datasets
  • Move from offline to online learning
  • Rank instruction-level tasks by value

Application scenarios

  • Search for an Italian restaurant on a phone
  • Find a listed product page and run a search
  • Hold steady performance through site changes

Pros & cons

✓ What we like

  • Offline to online learning uses trial and error
  • Self-collected rollouts reduce recovery failures

! What to watch out for

  • A software environment must be installed
  • Experiments need datasets and parameter setup

FAQ

What does the evaluator do?

The material says an autonomous VLM evaluator handles open-ended Android tasks.

How are tasks prioritized?

An instruction-level value function builds the automatic curriculum.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review