LIVE
OpenAI moves toward a rigorous lattice definition of the Standard Model15/11/26 · OpenAI|Open-Weight Reasoning Model Learns to Operate Real Robots Through Simulation-Only Training13/11/26|Forensic analysis finds traces of export-controlled GPUs in Chinese frontier models12/11/26|Google DeepMind unveils Gemini 4 Hydrogen12/11/26 · Google DeepMind|AlphaGenome 2: DeepMind refines its model for decoding non-coding DNA11/11/26 · Google DeepMind|Google DeepMind rolls out Gemini 4 Carbon10/11/26 · Google DeepMind|Google DeepMind unveils AlphaProof Code for formal software verification10/11/26 · Google DeepMind|OpenAI launches a dedicated business presence feature in ChatGPT10/11/26 · OpenAI|Anthropic launches investor roadshow ahead of planned IPO10/11/26 · Anthropic|A petition calls for Jensen Huang's return as Nvidia CEO16/11/26 · Nvidia|OpenAI strengthens tools against online child sexual exploitation16/11/26 · OpenAI|OpenAI Expands Free ChatGPT Access to 200,000 Researchers16/11/26 · OpenAI|OpenAI moves toward a rigorous lattice definition of the Standard Model15/11/26 · OpenAI|Open-Weight Reasoning Model Learns to Operate Real Robots Through Simulation-Only Training13/11/26|Forensic analysis finds traces of export-controlled GPUs in Chinese frontier models12/11/26|Google DeepMind unveils Gemini 4 Hydrogen12/11/26 · Google DeepMind|AlphaGenome 2: DeepMind refines its model for decoding non-coding DNA11/11/26 · Google DeepMind|Google DeepMind rolls out Gemini 4 Carbon10/11/26 · Google DeepMind|Google DeepMind unveils AlphaProof Code for formal software verification10/11/26 · Google DeepMind|OpenAI launches a dedicated business presence feature in ChatGPT10/11/26 · OpenAI|Anthropic launches investor roadshow ahead of planned IPO10/11/26 · Anthropic|A petition calls for Jensen Huang's return as Nvidia CEO16/11/26 · Nvidia|OpenAI strengthens tools against online child sexual exploitation16/11/26 · OpenAI|OpenAI Expands Free ChatGPT Access to 200,000 Researchers16/11/26 · OpenAI|
Research

Open-Weight Reasoning Model Learns to Operate Real Robots Through Simulation-Only Training

Researchers trained a 27-billion-parameter open-weight reasoning model entirely in simulation, with no human demonstrations, to directly control real-world robots, narrowing the gap with proprietary frontier systems.

November 13, 20264 min readPublished byarXiv

Robot-use agents designed for general-purpose control have so far delivered strong results mainly through proprietary frontier models, leaving open-weight alternatives trailing well behind. A new study narrows that gap by showing that an open-weight 27-billion-parameter reasoning model can learn to operate real robots natively, trained entirely through reinforcement learning in simulation and without any demonstration data.

The researchers built a dedicated post-training system called PEAR (Post-training Engine for Agentic Robotics) to run what they describe as robot-use reinforcement learning, or RURL. The model observes its environment, reasons through a chain of thought, and directly outputs the next end-effector pose, with no specialized software harness involved — an approach that mirrors how computer-use agents operate. Training took place entirely on the simulated LIBERO and RoboDojo benchmarks.

The resulting skills generalize well beyond the training setup. They transfer to a robot the model has never encountered, to the physical world, and to instructions it was never explicitly trained on. On a real dual-arm robot, the trained model succeeds in 60% of trials, compared with 0% for the untrained base model. On LIBERO, the success rate climbs from 1.6% to 74.4%, approaching the performance of the proprietary reference models cited in the paper. Against π0.5, a vision-language-action model fine-tuned on demonstrations, the new approach proves considerably more robust to task variation: on LIBERO-Pro, it succeeds in 57.4% of episodes involving unfamiliar tasks in familiar scenes, versus just 0.75% for π0.5.

Importantly, this specialization comes with no measurable cost to the model's general abilities, with performance holding steady across nine reasoning, coding, and vision benchmarks. The authors argue that robot-use reinforcement learning could become a standard component of foundation-model post-training, alongside computer-use and web-browsing skills — potentially reducing reliance on closed, proprietary systems for robotic control.

Tags
roboticsreinforcement-learningopen-weightreasoning-modelssimulationagents

Read also

Open-Weight Reasoning Model Learns to Operate Real Robots Through Simulation-Only Training · nAIvigate