27B reasoning model learns real-robot control from simulation alone
Original title: Training Reasoning Models in Simulation to Natively Use Real Robots
It succeeded in 60% of real dual-arm trials after simulation-only training; the untuned base model succeeded in none.
Researchers trained an open-weight 27B reasoning model with robot-use reinforcement learning entirely in simulation, using a Post-training Engine for Agentic Robotics and no demonstrations. The model outputs the next end-effector pose from observations and chain-of-thought, without a harness. Skills transferred to an unseen robot, the real world and novel instructions. On LIBERO, success rose from 1.6% to 74.4%, near GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort, with no measured drop on nine general benchmarks.
- After LIBERO and RoboDojo training, real dual-arm trials succeeded 60% of the time.
- On LIBERO-Pro, the model succeeded in 57.4% of episodes with a new task in a familiar scene.
- Demonstration-tuned π0.5 succeeded in 0.75% of those LIBERO-Pro episodes.
- The authors report no measurable degradation on nine reasoning, coding and vision benchmarks.
- Products
- PEARLIBERORoboDojoLIBERO-Pro
From the original article
1 min read at arxiv.orgRobot-use agents (RUAs) have emerged as a promising approach to general-purpose robot control. However, strong performance has so far been achieved only by frontier proprietary models such as GPT-6 Astra and Claude Fable 5. 2, whereas open-weight models remain substantially less capable.
By Jiaming Tang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.