A team at MIT led by Song Han, together with collaborators at UC Berkeley, UC San Diego and NVIDIA, has released PEAR, an open-source post-training engine for robotics with an unusually strong claim: it turns a general-purpose open-weight language model into a working robot controller without a single human demonstration and without any real-robot training data. All learning happens as reinforcement learning inside physics simulators. The work, described in a paper by Jiaming Tang, Zhuoyang Zhang, Junxian Guo and colleagues, is public on the lab's GitHub alongside the training stack.
The system, built on the open Qwen3.8-27B model, follows what the authors call a robot-use pattern that mirrors how computer-use agents operate. At each step the model receives a camera image and the robot's joint states, reasons through a chain of thought, and outputs a target pose for the gripper, which a low-level controller executes. There is no separate object detector and no grasp planner; perception, planning and error judgment all live in the same model, and each round's observation returns to the context for the next decision.
The headline numbers, all team-reported and not independently replicated: on the LIBERO benchmark, the base Qwen3.8-27B scores 1.6 percent zero-shot, which rises to 74.4 percent after PEAR's simulation-only reinforcement training, close to the 76.2 percent the team cites for GPT-6 Astra Max. On hardware, the trained model was taken directly to a YAM dual-arm rig it had never seen and completed 18 of 30 trials, a 60 percent success rate, while the untrained base model completed none.
Robustness is where the approach most clearly separates from imitation learning. Under the perturbed LIBERO-Pro setting, where environments, layouts and instructions shift, the established vision-language-action model pi-0.5 collapses from 96.9 percent to 0.75 percent, while PEAR retains 57.4 percent, apparently because it leans on the base model's general understanding instead of memorized trajectories.
Just as notable is what the training does not break. Across nine general benchmarks including AIME mathematics, GPQA science and LiveCodeBench, the robot-post-trained model scores essentially the same as the original Qwen3.8-27B, meaning one model still answers questions, writes code and controls a robot arm. The team also reports more than 220 tokens per second decoding on a single RTX 5090, which keeps inference local, a practical precondition for factory use where cloud round-trips add latency and data-compliance burdens.
The results land one week after OpenAI dissolved its own independent robotics team on the argument that general frontier models such as GPT-6 Astra can already control robots directly. PEAR is effectively an open-weight, single-GPU experimental version of that thesis, and it argues something stronger: post-training a general model in simulation can substitute for both dedicated robot foundation models and the expensive teleoperation data they are usually trained on, a data moat that much of the embodied AI industry is built on.
The caveats are real. Sixty percent success is far from the 99-percent-plus reliability industrial deployment expects; 30 trials is a small sample; and the demonstrated task range covers grasping, placing and simple tool use rather than assembly or deformable objects. The benchmarks are the authors' own. But if the success rates hold at scale, the expensive link in embodied AI, one operator and one robot collecting minutes of usable data per day, starts to look replaceable by simulation, and the debate over whether robots need their own foundation models gains its most concrete evidence point yet.
Comments (0)
Log in to join the discussion
Log InNo comments yet