An open model learns to work a robot arm entirely in simulation
Researchers trained a 27-billion-parameter open-weight reasoning model to control robot arms using only simulated practice. On a real two-armed robot it had never seen, it succeeded in 18 of 30 trials. The work is a preprint and has not been peer reviewed.

AI illustrationWhat the study shows
The team, from MIT, UC Berkeley, UCSD, NVIDIA and Inco AI, started from Qwen3.8-27B. This is an open-weight model, meaning anyone can download and run it. It can read images and reason in text before answering.
They set it up as a robot-use agent. This works like an AI that operates a computer from screenshots. In each round, the model looks at camera images and the arm's position, reasons in text, and outputs where the gripper should go next. Standard robot software then moves the arm there.
The model received no helper tools: no object detector, no motion planner, no library of ready-made skills. It had to find objects, plan moves and notice failures by itself.
Training used reinforcement learning, where a model learns from a reward signal. Here the reward was simply whether the simulated task succeeded. Training ran on two simulated benchmarks: LIBERO, with a single arm, then two tasks in RoboDojo, with two arms. It took about 50,000 GPU-hours.
On LIBERO, success rose from under 2% to 74.4%. The authors report this is close to the best closed models from OpenAI at their maximum reasoning setting, which scored about 76% and 81%.
The authors compared it with π0.5, a robot model trained by copying human demonstrations. π0.5 scored higher on the original LIBERO tasks. But when the scene stayed the same and the instruction asked for a different task, π0.5 succeeded under 1% of the time. The new model succeeded 57% of the time. The π0.5 figures come from earlier papers, not new runs.
On a real two-armed robot that no model had trained on, the final model succeeded in 18 of 30 trials. The base model succeeded in none. Scores on nine general tests of maths, coding and vision showed no measurable loss after robot training.
Our reading
In our reading, the main result is that one recipe for training computer-using AI also works for robots. The model never saw a real robot or a human demonstration. Yet it moved from failing almost every task to succeeding in most simulated ones, and then carried some of that skill to real hardware.
The comparison with π0.5 is the most interesting part. Models that learn by imitation tend to repeat the motion they were shown. This model reads the instruction and acts on it. In practice, that could mean a warehouse arm that can be told to pick a different item from a familiar shelf without fresh demonstrations. That remains a hope, not something this paper tested.
The real-robot evidence is thin. Thirty trials on two tasks is a lab check, not proof of reliability. The final model had also practised both tasks in simulation, on a different robot. We rate the work at readiness level III, open source available: the training engine is public on GitHub, but the paper does not say whether the trained model weights are released.
Read the rest of this story
Membership is free. You get every article in full and the Issue every Sunday.
We send a sign-in link, no password needed. Already a member? Sign inReadiness assessment
| Level | Status | Evidence |
|---|---|---|
| I Lab only | Met | Working method with training, simulation results and a small real-robot test on a dual-arm YAM. |
| II Benchmarked | Met | Results on public benchmarks: LIBERO 74.4%, LIBERO-Plus, LIBERO-Pro 62.1%, RoboDojo, and RoboLab-120 37.3% against leaderboard policies. Caveat: baselines such as π0.5 are copied from prior work. |
| III Open source | Met | The paper links https://github.com/mit-han-lab/pear for the PEAR training engine, and the base model Qwen3.8-27B is open-weight. We could not confirm from the paper text whether the trained Qwen3.8-27B-PEAR weights are released. |
| IV In production | Not yet | No documented use in production by a named organisation. The real-robot evidence is a lab test of 30 trials per model. |
What would move it up
Level 4 would need documented deployment by a named organisation outside the authors' lab. Useful steps: larger real-robot tests on tasks the model never trained on, independent replication, and a faster, shorter-reasoning model that handles long tasks without exhausting its context. Confirmed public release of the trained weights would firm up level 3.
The study
Jiaming Tang, Zhuoyang Zhang, Junxian Guo et al.. Training Reasoning Models in Simulation to Natively Use Real Robots. arXiv, 13 November 2026. Read the paper
Notes
- Tang, Zhang, Guo et al., Training Reasoning Models in Simulation to Natively Use Real Robots. arXiv preprint 2611.18562, November 2026. Not peer reviewed. Source
- PEAR training engine code, as linked from the paper. Source
- LIBERO average success: base model 1.6%, trained model 74.4%; GPT-6.1 Sol 76.2% and GPT-6 Astra 80.6% at max effort; 48.1% and 49.3% at medium effort (Figure 3b, Table 2).
- LIBERO-Pro, new task in a familiar scene: π0.5 0.75%, trained model 57.4%. Mean over four perturbations: 53.3% against 62.1% (Table 3). π0.5 on LIBERO: 96.9%.
- Real dual-arm YAM robot: final model 18/30, LIBERO-only model 7/30, base model 0/30 (Table 7).
- Full RoboDojo benchmark: final model 5.50% average; 370 of 540 episodes (68.5%) ended when the 256k-token context ran out (Appendix E).
- Speed: about 4.3 s per round of xhigh reasoning at 747 tokens/s on one GB300 GPU, using compressed NVFP4 weights. LIBERO success at low effort 21.9% (Figure 9). Training about 50,000 H100-equivalent GPU-hours (Appendix A).
- Readiness: level III, open source available. Training code is public; release of trained weights not confirmed in the paper.
Edited and signed by
Stefan PerdalEditor and publisher86 stories editedStefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.
How this was made
An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.
How we workNo corrections


