Skip to content
Trending storyDeveloping

Ai2 reports mixed ToyBlocks results for robot models

2 articles1 sourcesince Nov 6Last article 1h ago ·

Overview

AISummary of 2 articles

Ai2 says ManiFlow, π0.5 and MolmoAct 2 generalized poorly to new spatial configurations in its initial ToyBlocks tests.

The company says controlled experiments will test training and model design changes to find which techniques improve generalization.

In a separate report, Ai2 says Molmo 2 and ManiFlow, trained only in the ToyBlocks virtual world, spelled a word with lettered blocks on a physical robot in 15 of 30 attempts. In simulation, the same models succeeded in 111 of 200 attempts. Ai2 says more trials are needed to confirm that simulation-trained models transfer to physical robots.

Written by AI from the articles below · updated Nov 6, 9:22 AM ET

Check the sources:

Developments

2 developments

  1. Nov 6, 9:13 AM ET · 1 article
    Ai2 reports simulation-to-robot transfer experiment
    Ai2: Ai2 reports simulation-trained Molmo 2 and ManiFlow transfer to physical robot
  2. Nov 6, 9:13 AM ET · 1 article
    Ai2 reports initial ToyBlocks generalization tests
    Ai2: Ai2 finds ManiFlow, π0.5, and MolmoAct 2 generalize poorly in ToyBlocks tests

Article timeline

The articles in this story. Times are ET.

Nov 6
  1. Ai2Official
    Ai2 reports simulation-trained Molmo 2 and ManiFlow transfer to physical robot

    AIAi2 says Molmo 2 and ManiFlow, trained only in the ToyBlocks virtual world, spelled a word with lettered blocks on a physical robot in 15 of 30 attempts. In simulation, the same models succeeded in 111 of 200 attempts. Ai2 says simulation-trained models can transfer to physical robots, though more trials are needed.

    Image from @allen_ai's post
  2. Ai2Official
    Ai2 finds ManiFlow, π0.5, and MolmoAct 2 generalize poorly in ToyBlocks tests

    AIAi2 says ManiFlow, π0.5, and MolmoAct 2 generalized poorly to new spatial configurations in its initial ToyBlocks tests. The company says controlled experiments will test changes to training and model design to measure which techniques improve generalization.

    Image from @allen_ai's post

Heat Trend

Not enough continuous observations to show a trend yet.