A new paper introduces Zero-real Agent Policy (ZAP), a method for teaching robots to perform manipulation tasks without ever training on real-world robot data. The approach starts from a pretrained 27-billion-parameter vision-language model, which already carries general world understanding and reasoning capabilities. The researchers then apply supervised fine-tuning and reinforcement learning (GRPO) on simple simulated tasks to teach the model how to actually move a robotic arm, using a lightweight tool harness to connect the model's outputs to physical actions.
The resulting system reportedly reaches 96.7% success on real-world evaluation tasks, a figure the authors say matches a reference system called GPT-6 Astra tested under the same harness, while substantially outperforming a comparison vision-language-action model, MolmoAct2, which scored 33.3%. ZAP is also described as three times faster and seventeen times cheaper to run than the GPT-6 Astra baseline, which, if confirmed, would make it a notably more practical option for deployment.
Beyond raw success rates, the paper emphasizes generalization. ZAP is shown handling tasks and objects absent from its training data: placing a test tube into a specific hole of a rack on request, passing a test tube between two robotic arms in mid-air, marking an X on a tic-tac-toe board, and playing chess. None of these scenarios were part of the simulated training curriculum, suggesting the model's pretrained reasoning transfers into novel physical and linguistic contexts rather than simply memorizing trained behaviors.
The system also transfers to a robot embodiment it was never trained on. Tested on a single-arm Franka robot, ZAP achieves 55.9% success on the RoboLab-120 benchmark, edging out the best competing vision-language-action or world-action-model baseline by 4.9 percentage points. This cross-embodiment transfer is one of the more demanding tests for generalist robot policies, since control strategies often do not carry over cleanly between different hardware setups.
The core claim worth noting is architectural: separating 'what' a task requires, handled by a large pretrained model, from 'how' to execute it physically, learned cheaply in simulation, may offer a more data-efficient path to generalist robot control than collecting large real-world robot datasets. If the reported numbers hold up under independent scrutiny, this split-training strategy could reduce the cost and time needed to deploy capable robot policies across new tasks and hardware.