← Back to the feed
Robotics & manufacturingResearch findings

Training Open-Weight Reasoning Models to Natively Use Real Robots in Simulation

Original published 2026.11.14 03:59 KSTAdded to Blackdeck 2026.11.17 08:13 KST

What has changed?

An open-weight 27B reasoning model is trained to natively use real robots through simulation-based reinforcement learning.

Key points

  1. An open-weight 27B reasoning model is trained to use real robots directly via simulation-based reinforcement learning.
  2. The model achieves 60% success rate on a real dual-arm robot after simulation training.
  3. It generalizes well to new environments and instruction variations.

Connect it to your work

Skills learned in simulation generalize well to unseen robots and novel instructions.

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. arxivTraining Reasoning Models in Simulation to Natively Use Real Robots