← Back to the feed
Robotics & manufacturingResearch findings

Pretraining Knows What, Simulation Teaches How

Original published 2026.10.27 02:44 KSTAdded to Blackdeck 2026.11.02 14:12 KST

What has changed?

A new method, ZAP, uses pre-trained vision-language models to learn robot control through simulation without real-world data.

Key points

  1. A robot policy must understand what a task requires in a scene and how to move to achieve it.
  2. ZAP teaches a small vision-language model to control a robot via simulation using a simple tool harness.
  3. Generalizes to completely new objects, goals, and manipulation strategies not seen during training.

Connect it to your work

Generalizes to unseen objects, goals, and manipulation strategies.

What remains unproven

This research note is based on an abstract; check the original for test conditions and application scope.

Originals used for this note

Links open the original publisher in a new tab.

  1. arxivPretraining Knows What, Simulation Teaches How: Compositional Generalization in a Sim-Trained Robot Agent