Wakeline

Research

Learning in Worlds

Learning in Worlds is our playground for the questions behind general intelligence. What does a learner need when nobody explains the world? How do words come to mean something? We build small worlds where things change, give our learner a body and senses, and watch it figure things out.

The worlds

An overhead view of the arena world: the two-antenna creature moves among green food circles, pink hazard diamonds and dark circular obstacles. Blue, amber and violet glows distinguish chemical channels.

Arena

What happens when the world changes and nobody says so?

A small creature lives one continuous life in a 2D world of smells, food and hazards. It senses 13 readings, gets no reward labels and has to notice for itself when things change.

What changes

  • Novel food
  • Meaning flip
  • Sensor damage
  • Obstacles
  • Fast switching

See studies in the Arena

Illustration: a small grid world with coloured shapes, and an instruction that points to one of them. walk to the small violet circle

Grounded language

What does it take for words to mean something in a world?

Here, instructions have to be connected to what is happening in a small world, and familiar words combined in new ways. We are exploring this with the gSCAN and ReaSCAN benchmarks.

See studies in Grounded language

Studies

3 studies

DateTitleWorldWhat it exploresStatus
A world that doesn't hold still Arena What happens when the world changes and nobody says so? Published
Language as a second view of a learned world Grounded language Can a learner find objects and words in raw pixels and sound, and then use the words to act? Published
Learning to follow commands from demonstrations Grounded language Can words get their meaning from the world they describe? Published