Qwen3.8-27B Completes 18 of 30 Trials on a New Robot
The model practiced pickup and block alignment in two simulators before testing on a dual-arm YAM robot. It succeeded in 12 of 20 pickup attempts and six of 10 alignment attempts.

Qwen3.8-27B completed 18 of 30 tests on a dual-arm YAM robot after reinforcement learning in simulation, showing that robot skills transferred to unfamiliar hardware for two practiced tasks.
The research team used the LIBERO and RoboDojo simulators to post-train Qwen3.8-27B, an open-weight reasoning model with 27 billion parameters. The robot-use training used neither demonstrations nor data from physical robots. The team submitted its preprint Friday, Nov. 13, at 1:59:02 p.m. ET (18:59:02 UTC).
The work describes a robot-use post-training stage; it does not describe how the underlying model was originally built. For the physical evaluation, the researchers used a dual-arm YAM robot, hardware absent from the model’s simulation practice. The final model had practiced pickup and block alignment in RoboDojo on different equipment.
The final model completed 12 of 20 pickup attempts and six of 10 block-alignment attempts. The base model completed zero of 30 trials, while a version trained only in LIBERO completed seven. All three versions followed the same YAM protocol with xhigh reasoning effort.
Pickup trials required raising one of 10 objects by 10 centimeters; alignment trials used five block layouts. The model processed camera views and robot-state readings, then returned a target pose for its end effector. Inverse-kinematics software and the robot controller converted that target into movement.
Successful runs by the final model lasted seven to 24 interaction rounds and used 23,000 to 132,000 context tokens. A trial ended after 25 rounds or if any round exceeded 32,000 reasoning tokens. The PEAR authors wrote, “Our model is still slow at carrying out tasks, because it reasons at length before every move.”
The experiment measures transfer to unfamiliar hardware on a narrow set of practiced tasks; it does not establish performance on unrelated jobs. The results cover 30 trials, two task types and one physical robot. The work adds to research on simulated robot training, including NVIDIA’s work with GR00T 2 and Isaac Lab 3.0.