Models
A preprint with MIT authors says Qwen3.8-27B-PEAR, trained only in simulation, completed 18 of 30 trials on a real dual-arm robot it never trained on, while the base model completed none
A preprint says a 27B model trained only in simulation completed 18 of 30 real-robot trials; its base model completed none. The tests covered two task types.
A preprint dated November 13, 2026, with authors from MIT, UC Berkeley, UCSD, Nvidia and Inco AI, reports that an open-weight 27-billion-parameter reasoning model learned to control real robots after training only in simulation. The training system is called PEAR, for Post-training Engine for Agentic Robotics, and the resulting model is Qwen3.8-27B-PEAR. On a dual-arm YAM robot that no model was trained on, the final model succeeded in 18 of 30 trials. These are trials of a small task set, not 30 different tests.
The essentials
- What PEAR is: a post-training system that applies reinforcement learning to a reasoning model in simulation, without human demonstrations. The model is Alibaba’s open-weight Qwen3.8-27B.
- Training: LIBERO and RoboDojo simulators, 148 steps on three LIBERO suites, then 77 more after adding LIBERO-Long, then 115 on RoboDojo. No data from real robots was used.
- Real-robot result: 18 of 30 trials for the final model, 7 for a LIBERO-only checkpoint and 0 for the base model, on a dual-arm YAM it had never seen.
- What the 30 trials were: pickup (12 of 20 succeeded) and block alignment (6 of 10).
Trials, not 30 different tests
The headline figure counts attempts at two task families, pickup and block alignment, on one robot. It shows transfer from simulation to a new robot within those families. It does not show that the model can do 30 different jobs, and the paper’s two prompted reasoning demos are not scored.
What the authors flag as limits
The paper reports no component ablations, one base model, two simulators and a small task set. Results depend on the prompt and on the controller reaching the target poses. Reasoning is slow, about 4.3 seconds per high-effort round. The authors also say the model needs the robot and its coordinate directions described in the prompt.
What is not established
- Whether the result holds on other robots or tasks: not tested beyond one YAM and two task families.
- Which component of the recipe matters most: no ablations reported.
- Peer review: this is a preprint.
- Real-world speed and reliability: reasoning takes seconds per step.
- Independent replication: none reported.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Paper | Preprint of November 13, 2026, arXiv 2611.18562 | arXiv |
| Model | Qwen3.8-27B post-trained with PEAR by reinforcement learning | arXiv |
| Training data | Simulation only, no demonstrations, no real-robot data | ArXivSignals |
| Real robot | Dual-arm YAM, unseen in training; 30 trials | ArXivSignals |
| Results | Final 18/30, LIBERO-only 7/30, base 0/30 | Sam on AI |
FAQ
Is PEAR a model or a method?
PEAR is the training system. The model it produces is called Qwen3.8-27B-PEAR.
Did it pass 18 of 30 different tests?
No. It succeeded in 18 of 30 trials across two task families, pickup and block alignment.
Was any real-robot data used?
The paper says no. Training used simulation only, without demonstrations.
Method and limits
Checked on November 16, 2026. We read the arXiv paper and two summaries through search excerpts; one direct fetch of the paper was rate-limited. The request said ‘18 of 30 tests’; the paper counts trials on two task families. Affiliations come from the arXiv PDF excerpt; one summary page lists only lab tags.
Sources
- arXiv: Training Reasoning Models in Simulation to Natively Use Real Robots (HTML) (verified 2026-11-08)
- arXiv: Training Reasoning Models in Simulation to Natively Use Real Robots (PDF) (verified 2026-11-08)
- ArXivSignals: Training Reasoning Models in Simulation to Natively Use Real Robots (verified 2026-11-08)
- Sam on AI: AI Brief - November 16, 2026 (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy