An open-weight 27-billion-parameter reasoning model learned robot skills in simulation and then completed 18 of 30 trials on a real dual-arm robot it had never trained on. The approach keeps inverse kinematics and motion control outside the language model, letting the same pose-based action interface move from simulated single-arm and dual-arm platforms to hardware without real-robot fine-tuning.
What Did the Researchers Build?
The team trained a reasoning model to use robots through reinforcement learning in simulation. Rather than having the model directly control motors, the system asks it to choose a target pose for the robot’s end effector—the point at the end of the arm, such as a gripper. A separate robot controller turns that target into the joint movements needed to reach it.
This separation creates a shared action interface across different machines. The researchers used simulated tasks with a single-arm Franka robot and a dual-arm ARX robot, then deployed the trained model on a real dual-arm YAM robot. For each machine, the prompt describes the robot’s parts, coordinate frame, and cameras. The robot’s existing inverse-kinematics software and controller handle the physical motion.
The model started from Qwen3.8-27B, an open-weight reasoning model. Training used simulated task outcomes rather than demonstrations of people operating robots. The researchers’ central test was whether skills learned in simulation could transfer not only to tasks the model had not practiced, but also to a different simulator and a real robot. They report that deployment on the real robot required no additional training on real-robot data.

What Results Did the Model Achieve?
The strongest result was zero-shot deployment on the real YAM: the final model succeeded in 18 of 30 trials, while the base model succeeded in none. A model checkpoint trained only on LIBERO simulation succeeded in 7 of 30 real-robot trials, despite having seen neither those tasks nor that robot. “Zero-shot” here means the model was tested without additional training on the new robot.
The model also transferred skills across simulation environments. After training only on LIBERO, it achieved 39.9% success on RoboDojo’s general-pickup task and 17.0% on align-blocks, compared with 10.7% and 13.0% for the base model. Training on additional RoboDojo tasks brought the reported success rate across those tasks to 75.0%.
| Evaluation | Base model | Trained model | What changed |
|---|---|---|---|
| Real dual-arm YAM trials | 0/30 | 18/30 | Final model tested without real-robot training |
| Real YAM trials, LIBERO-only checkpoint | — | 7/30 | Transfer from simulation to unseen hardware |
| RoboDojo general pickup, LIBERO-only training | 10.7% | 39.9% | Transfer to a different simulator and robot |
| RoboDojo align blocks, LIBERO-only training | 13.0% | 17.0% | Transfer to a different simulator and robot |
| LIBERO-Long without direct training | 0.4% | 39.0% | Transfer from other LIBERO task suites |
| LIBERO-Long after adding it to training | — | 53.2% | Training provides direct experience with the suite |
The gains did not come with a large drop in the model’s other measured abilities. It scored 98.5 on AIME, 89.7 on GPQA Diamond, and 88.4 on LiveCodeBench, while staying within two points of the base model on every vision benchmark tested.
How Does the Simulation-to-Robot System Work?
The model receives a prompt describing the task, the robot’s embodiment, and its camera views. It then reasons about what to do and outputs a target pose for the end effector. The target is an instruction such as where and how the gripper should be positioned—not a stream of commands to individual motors. A separate inverse-kinematics system calculates joint positions, and the robot controller moves the arm to the target.
That interface is key to transferring between robots. The action format stays the same, while the prompt changes to describe the robot’s parts and coordinate frame. The controller also stays outside the model. In principle, this lets the same trained policy work with different robot hardware as long as the robot can accept pose targets and has a suitable controller.
Training uses reinforcement learning with a binary outcome: an episode either succeeds or fails. At the start, successful episodes are rare, so many tasks provide little useful learning signal. To address this, the researchers used a curriculum. Training began with LIBERO’s Goal, Spatial, and Object task suites; later it included all four LIBERO suites, followed by two RoboDojo tasks. Skills from the easier or more learnable tasks helped the model begin succeeding on harder tasks, creating more examples from which reinforcement learning could improve it.
The approach has a deliberate trade-off: the model chooses one pose at a time rather than a multi-step motion sequence. That keeps the interface simple, but a complex action can require several rounds of reasoning and movement.

Why Does This Matter for Robotics?
Robots are often expensive and time-consuming to train directly because every failed attempt on physical hardware can consume operator time, equipment capacity, and sometimes replacement parts. This work tests a route in which the learning happens in simulation, while an existing controller translates the model’s high-level decisions into robot motion. The real-robot result suggests that simulation training can produce useful behavior on hardware the model has not encountered during training.
The shared pose-target interface could also make it easier to reuse a robot-use model across different platforms. A buyer or engineering team would still need to provide accurate robot descriptions, camera information, and reliable control software, but the model would not necessarily need a new training run for every robot. That is a practical distinction for organizations evaluating fleets with varied hardware.
The evidence is an early demonstration, not a ready-made deployment recipe. Real-world trials were limited, performance depended on detailed prompts, and each action required reasoning before the controller moved the arm. Still, the work offers a useful reference point for teams comparing robot platforms and planning simulation-led development. Teams evaluating hardware can also browse used robots for sale.
What Are the Limitations and Open Questions?
The training covered only two simulators and a small set of tasks, including just two RoboDojo tasks. The researchers stopped scaling largely because of limited GPU resources, and the model started from a 27-billion-parameter base. The results do not yet show how well the approach handles a broad range of robots, environments, or everyday operational conditions.
The system is also slow at task execution. A reasoning round takes about 4.3 seconds on average, and some reasoning sequences reach 32,000 tokens, making a single move take more than half a minute. The model acts one pose at a time, so complex movements require extra rounds. Its success also depends heavily on prompts that clearly explain the robot and task. Finally, controller quality matters: inaccurate motion can prevent the arm from reaching the model’s target pose.
What Are the Frequently Asked Questions?
Did the model train on real-robot data? No. The reported real-robot tests used a model trained in simulation, without additional training on real-robot data.
What does the model output? It outputs a target pose for the robot’s end effector. A separate inverse-kinematics system and controller convert that target into robot movement.
Did the model work on a robot it had never trained on? Yes. The final model was tested on a real dual-arm YAM that was not part of its training, succeeding in 18 of 30 trials.
What is the main trade-off? The model can transfer between simulated and real robots without directly controlling their motors, but it reasons before each pose-level action, which can make tasks slow.
What Is the Practical Takeaway?
Simulation-trained reasoning models can transfer useful robot skills to unfamiliar tasks, simulators, and real hardware when paired with a consistent pose-based interface. The results are promising, but faster action, broader task coverage, and less prompt dependence remain important steps toward practical deployment.
