Researchers train AI in simulation and unlock real robot skills
Quick Take
- PEAR researchers report 18 successes in 30 physical trials after simulated robot training, compared with none for the base model.
- The final model practiced pickup and block alignment in simulation before testing those task types on different physical hardware.
- Lower-precision decoding speeds describe a separate configuration from the evaluations.
Researchers behind PEAR, a system for training AI models to control robots, reported that a model completed 18 of 30 physical trials after learning its robot skills entirely in simulation.
The preprint submitted Nov. 13 describes robot post-training of Qwen3.8-27B, an open-weight reasoning model with 27 billion parameters. Researchers used reinforcement learning without demonstrations during that stage.
The result offers evidence that simulated practice can teach skills that carry over to different physical hardware. The final model had already practiced the two evaluated task types, pickup and block alignment, in simulation.
What transferred to the real robot
The physical test used a dual-arm YAM robot, which none of the compared models received training on in this study. It included 20 pickup attempts across ten objects and ten alignment attempts across five block layouts, with each object or layout tried twice.
The final model succeeded in 12 pickup attempts and six alignment attempts. A checkpoint trained only on the LIBERO simulation benchmark completed seven of the 30 trials.
The final model had continued training on the corresponding two tasks in RoboDojo using a different robot, according to the paper's physical-trial account.
Practice in simulation. Test on real hardware.
PEAR researchers report 18 successes in 30 physical trials after robot post-training in simulation. Explore how the tested checkpoints compare.
Compare the checkpoints on the real robot
Final trained model: 18 of 30 physical trials succeeded. LIBERO, then two RoboDojo task types.
Two practiced task types transferred to a different physical robot; the final model completed 18 of 30 trials.
Author-reported preprint, November 13, 2026; not independent replication. The final model practiced both evaluated task types in simulation. No real-robot training data were used in this study; simulation-only describes robot post-training, not foundation training. Prompt dependence and long reasoning remain limits.
Sources
Supporting facts
- PEAR researchers report training Qwen3.8-27B, an open-weight 27-billion-parameter reasoning model, with robot-use reinforcement learning entirely in simulation and without demonstrations during this robot post-training. [1] [3]
- In the authors' physical experiment, the final model completed 18 of 30 trials on a dual-arm YAM robot, the LIBERO-only checkpoint completed 7 of 30, and the base model completed none; all used the same protocol at xhigh reasoning effort. [1]
- The authors' 30-trial physical evaluation comprised 20 pickup attempts across ten objects and ten block-alignment attempts across five layouts; the final model succeeded in 12 pickup and six alignment trials. [1] [2]
- The final model had practised the corresponding pickup and block-alignment tasks on a different robot in RoboDojo simulation; none of the compared models was trained on the real YAM robot or real-robot data in this study. [1]
Method
Schematic training-to-evaluation path for the final model. Each dot represents one physical trial, grouped by aggregate outcome rather than actual trial order. Reported totals are 0/30, 7/30 and 18/30; no new rates or per-task checkpoint counts are derived.
Scope
Author-reported preprint, November 13, 2026; not independent replication. The final model practiced both evaluated task types in simulation. No real-robot training data were used in this study; simulation-only describes robot post-training, not foundation training. Prompt dependence and long reasoning remain limits.
CryptoSlateAt each turn, the model reads camera images and the robot's state, reasons in text, and specifies changes to the gripper's pose. Inverse kinematics translates the target pose into joint movements, which a controller executes.
The prompt supplies coordinate frames, gripper orientation, cameras, and task-completion conditions, and the authors acknowledge that performance depends on those descriptions.
All three models used the same high reasoning-effort setting, with trials having a 25-round budget and failing if a single round exceeded 32,000 reasoning tokens. The final model's successful attempts required 7 to 24 rounds.
The authors say reasoning before every move slows tasks. Their separate decoding demonstration reached 747 tokens per second on a GB300 using lower-precision NVFP4 weights and DFlash2, a decoding accelerator.
That suggests about 4.3 seconds to generate an average high-effort reasoning round. The reported evaluations used BF16 weights because quantization reduced accuracy, so the speed demonstration and physical success rate describe different operating conditions.
The public PEAR repository provides training infrastructure and deployment recipes. The result remains an author-reported experiment pending independent replication, and turning it into a broadly useful robot system will require wider task coverage and shorter reasoning delays.
Editorial credits
Also known as "Akiba," Liam Wright is a reporter, podcast producer, and Editor-in-Chief at CryptoSlate. He believes that decentralized technology has the potential to make…
Gino Matos is a law school graduate and a seasoned journalist with six years of experience in the crypto industry. His expertise primarily focuses on the Brazilian blockchain…