Wednesday, November 18, 2026

The PEAR research team, which developed a system for training AI models to control robots, reported that one model succeeded in 18 out of 30 physical trials after acquiring all of its robotic capabilities exclusively through simulation.

A preprint submitted on November 13 details post-training of Qwen3.8-27B, an open-weight reasoning model with 27 billion parameters, for robotic control. During this stage, researchers applied reinforcement learning without using human demonstrations.

The findings suggest that practice in simulation can impart skills that transfer to different physical hardware. Before real-world testing, the final model had rehearsed both evaluated task categories—object pickup and block alignment—entirely in simulated environments.

What transferred to the real robot

Physical evaluations were conducted on a dual-arm YAM robot, a platform that none of the comparison models had been trained on during this study. The test comprised 20 pickup attempts across ten objects and ten alignment attempts across five block configurations, with each object or configuration attempted twice.

The final model completed 12 successful pickups and six alignments. For comparison, a checkpoint trained solely on the LIBERO simulation benchmark succeeded in seven of the 30 trials.

According to the paper’s account of the physical trials, the final model continued training on the same two tasks within RoboDojo, though using a different robot.

ROBOT POST-TRAININGPHYSICAL EVALUATIONLIBEROSimulation practiceRoboDojoTwo task typesReinforcement learning changesthe reasoning model’s weightsTRANSFERDual-arm YAMDifferent physical robotPickup + block alignmentSame task types as RoboDojoNo physical demonstrations during this robot post-training.The study tests transfer of practiced task types to new hardware.ROBOT POST-TRAININGLIBERORoboDojoSimulation reinforcement learningchanges the model’s weightsTransferDual-arm YAMDifferent hardwarePickup + alignment practiced in simNo physical demonstrations duringthis robot post-training.Schematic transfer path for the final model.

Compare the checkpoints on the real robot

Compare the checkpoints on the real robotBase modelLIBERO onlyLIBERO + RoboDojo

Base model: 0 of 30 physical trials succeeded. No robot post-training in this study.

Base modelNo robot post-training in this study0/30successful physical trialsOne dot per trial; grouped by outcome, not trial order.Same YAM protocol: 20 pickup attempts + 10 alignment attempts, all at xhigh effort.SuccessfulUnsuccessful

Base modelNo robot post-training in this study0/30successful trialsOne dot per trial; grouped by outcome.20 pickup attempts + 10 alignment attemptsSuccessfulUnsuccessful

LIBERO-only checkpoint: 7 of 30 physical trials succeeded. Robot training in LIBERO simulation.

LIBERO-only checkpointRobot training in LIBERO simulation7/30successful physical trialsOne dot per trial; grouped by outcome, not trial order.Same YAM protocol: 20 pickup attempts + 10 alignment attempts, all at xhigh effort.SuccessfulUnsuccessful

LIBERO-only checkpointRobot training in LIBERO simulation7/30successful trialsOne dot per trial; grouped by outcome.20 pickup attempts + 10 alignment attemptsSuccessfulUnsuccessful

Final trained model: 18 of 30 physical trials succeeded. LIBERO, then two RoboDojo task types.

Final trained modelLIBERO, then two RoboDojo task types18/30successful physical trialsOne dot per trial; grouped by outcome, not trial order.Same YAM protocol: 20 pickup attempts + 10 alignment attempts, all at xhigh effort.SuccessfulUnsuccessful

Final trained modelLIBERO, then two RoboDojo task types18/30successful trialsOne dot per trial; grouped by outcome.20 pickup attempts + 10 alignment attemptsSuccessfulUnsuccessful

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.

During operation, the model processes camera images and the robot’s current state, reasons through the situation in text, and determines adjustments to the gripper’s position. Inverse kinematics then converts that target pose into joint movements executed by the controller.

The prompt provides coordinate frames, gripper orientation, camera inputs, and task-completion criteria. The authors note that performance depends heavily on these descriptions.

All three models operated under the same high reasoning-effort setting. Trials were limited to 25 rounds and failed if any single round exceeded 32,000 reasoning tokens. Successful attempts by the final model required between 7 and 24 rounds.

The authors observe that reasoning before every action slows down task execution. In a separate decoding demonstration, the system achieved 747 tokens per second on a GB300 using lower-precision NVFP4 weights and DFlash2, a decoding accelerator.

That throughput implies roughly 4.3 seconds per average high-effort reasoning round. Because quantization degraded accuracy, the published evaluations relied on BF16 weights; consequently, the speed demonstration and the physical success rate reflect different operating conditions.

The public PEAR repository offers training infrastructure and deployment recipes. The results remain an author-reported experiment awaiting independent replication, and developing the system into a broadly useful robot platform will require broader task coverage and reduced reasoning latency.

Source link

Exit mobile version