MIT Han Lab trains an open 27B reasoning model to control real robots in pure simulation — 1.6%→74.4% on LIBERO, 0%→60% on a robot it never touched
PEAR trains Qwen3.8-27B on binary task-success rewards in simulation with no demonstrations or harness; it transfers zero-shot to a real dual-arm robot and, unlike imitation-trained VLAs, keeps following instructions when the task changes.

MIT Han Lab trained an open-weight 27B reasoning model to control real robots, using reinforcement learning entirely in simulation — no demonstrations, no harness. It transferred zero-shot to a real dual-arm robot it never touched, going from 0% to 60% success. On LIBERO, training raised the model's success rate from 1.6% to 74.4%, and unlike a demonstration-trained vision-language-action baseline, it kept following instructions when the task changed.
What shipped
The team (MIT Han Lab with UC Berkeley, UCSD, NVIDIA, and Inco AI) built a Post-training Engine for Agentic Robotics (PEAR) and ran robot-use reinforcement learning (RURL) on Alibaba's Qwen3.8-27B, an open-weight, Apache-2.0, vision-capable reasoning model. The approach mirrors how frontier models acquire computer use: the model observes camera views, reasons in a chain of thought, and writes the next end-effector pose as text — no perception tools, planners, or skill libraries (paper, project blog).
Each round the model gets three camera views, the arm state, the instruction, and all prior rounds in context; it reasons in text and emits the next pose as JSON (e.g. {"left_arm": {"x": 0.0, "y": 1.0, "z": -17.4, ...}}). Inverse kinematics converts the pose to joint targets and the arm's controller executes it. The interface is identical in simulation and on the real robot.

Training runs on a binary reward — 1 if the simulator reports the task complete by episode end, otherwise 0 — using GRPO on the veRL stack, with a curriculum that adds LIBERO-Long and then two RoboDojo tasks after the LIBERO Goal/Spatial/Object suites. The paper reports LIBERO success climbing to 64.0% at step 148, 69.5% at step 184, and 74.4% at step 225 (blog, alphaXiv summary).
The trained weights, code, and training infrastructure are released (GitHub, Hugging Face).
Why it matters
The authors argue that chain-of-thought reasoning, not imitation, is what buys robustness. They compare against π0.5, Physical Intelligence's vision-language-action model fine-tuned on demonstrations, and the gap shows up exactly where imitation should fail: when the task changes.
On LIBERO-Pro, where the model is given a new task in a familiar scene, PEAR's model succeeds in 57.4% of episodes while π0.5 manages 0.75% — π0.5 falls 96.2 points from its 96.9% baseline, while PEAR falls only 17.0 from 74.4%. On swapped object positions it's 57.5% vs π0.5's 20.8%. The reasoning model keeps following the instruction; the imitation model keeps doing the motion it was shown (blog, alphaXiv).
Zero-shot transfer is the other headline. A model trained only on LIBERO reaches 39.0% on LIBERO-Long (never trained on, vs 0.4% base) and 39.9%/17.0% on RoboDojo's dual-arm ARX tasks. On RoboLab-120, the LIBERO+RoboDojo model scores 37.3% vs the π0.5 leaderboard entry's 28.0% (paper).
It also transfers to hardware it never saw. On a YAM dual-arm robot — not in training, no real-robot data — the final model succeeded in 18/30 trials (60%) vs 0/30 for the base model; the LIBERO-only checkpoint managed 7/30 (23%) (blog, alphaXiv).

The authors also report the model keeps its general capabilities, with no measurable degradation across nine reasoning, coding, and vision benchmarks (blog).
Caveats
Every benchmark number here is the authors' own, reported in a preprint that has not been independently replicated. Treat them as vendor claims.
- The reasoning-vs-imitation claim is not isolated by ablation. The systems differ in architecture and training recipe, so the robustness gap vs π0.5 can't be attributed to chain-of-thought reasoning alone — alphaXiv's analysis flags that no ablation isolates reasoning as the cause (alphaXiv).
- The real-robot evidence is thin: 30 trials, two tasks, and the painting and chemistry demos are unscored and depend on detailed prompts describing the embodiment and task (alphaXiv).
- The π0.5 comparison numbers were not re-run by the authors; they're taken from prior work (alphaXiv).
- Training compute is not disclosed in the accessible sources (the GitHub README lists hardware but no GPU-hours or FLOPs).
- The frontier-model comparison (GPT-6 Astra, GPT-6.1 Sol) is the authors' own evaluation, not independently confirmed.
The honest read: this is strong evidence that the recipe works and is worth building on, not a claim of reliable general-purpose deployment. The gap between the bold claim and the 30-trial real-robot evidence is the story to watch.