Number 9 of 12 · November 17, 2026

Training Reasoning Models in Simulation to Natively Use Real Robots

An open-weight 27B reasoning model, trained only in simulation with reinforcement learning and no demonstrations, reached 60% success on a real dual-arm robot where the base model reached none.

What they found

The authors train an open-weight 27B reasoning model to control robots the way a computer-use agent works with a screen. The model looks at the scene, thinks in a chain of thought, and outputs the next end-effector pose (where the robot's hand should go), with no extra harness. They built a training system called PEAR and a method called robot-use reinforcement learning (RURL), and trained on simulated tasks from LIBERO and RoboDojo. On a real dual-arm robot the trained model succeeded in 60% of trials, and the base model succeeded in none. On LIBERO, success rose from 1.6% to 74.4%, approaching GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort. On LIBERO-Pro, with a new task in a familiar scene, π_0.5 (fine-tuned on demonstrations) succeeded in 0.75% of episodes and the new model in 57.4%. The authors report no measurable loss on nine reasoning, coding and vision benchmarks.

Why it matters

Until now, the abstract says, strong robot control by general models came only from frontier proprietary systems. This work suggests an open-weight model can get there by practicing in simulation, without costly human demonstrations. The skills also carried over to an unseen robot, the real world and new instructions, which is a hard test for robot learning. The abstract does not say how many real-world trials were run or how many tasks they covered, so the 60% figure should be read with that in mind.

How it changes the race

If the results hold, groups without access to proprietary frontier models could build capable robot-controlling SI using open weights and simulation. It also points to robot-use training as one more step in post-training for general models.

Safety and control

The abstract says nothing about safety. A general reasoning model that controls real robots with no harness raises the usual questions about oversight and testing of physical actions. The paper's abstract does not address them.

The abstract, by the authors

Robot-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. We show that an open-weight 27B reasoning model can be trained to use real robots natively through reinforcement learning, entirely in simulation and without demonstrations. Like a computer-use agent (CUA), the model observes the environment, reasons in a chain of thought, and generates the next end-effector pose, without any harness. We build a Post-training Engine for Agentic Robotics (PEAR) and use it to train the model on simulated robot tasks at scale with robot-use reinforcement learning (RURL), mirroring how frontier models acquire computer use. We find that the skills acquired in simulation generalize well beyond the training tasks: they transfer to an unseen robot, to the real world, and to novel instructions. After training on LIBERO and RoboDojo in simulation, the model succeeds in 60% of our trials on a real dual-arm robot, whereas the base model succeeds in none. On LIBERO, training raises the success rate from 1.6% to 74.4%, approaching GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort. Compared with π_0.5, a vision-language-action model fine-tuned on demonstrations, our model is substantially more robust to changes in the task itself. On LIBERO-Pro, π_0.5 succeeds in only 0.75% of episodes when assigned a new task in a familiar scene whereas our model succeeds in 57.4%. Finally, the model retains its general capabilities, with no measurable degradation on nine reasoning, coding and vision benchmarks. These results highlight robot-use reinforcement learning as a promising component of foundation-model post-training.

Quoted from arXiv:2611.18562. The research and its words belong to its authors.

The authors

Jiaming Tang, Zhuoyang Zhang, Junxian Guo, Fangzhou Zhao, Yilong Zhao, Zekai Li, Xingyang Li, Yuyang Zhao, Song Bian, Zhijian Liu, Song Han

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.