Event
ZAP: Training Robot Policies Without Real Data
1 reports · 1 sources
Benjamin Pikus and co-authors propose Zero-real Agent Policy (ZAP): using a pretrained 27B-parameter vision-language model to control robots through a simple tool interface, then training it via supervised fine-tuning and GRPO on simple simulation tasks, with no real robot data used throughout training. The authors report that ZAP achieves a 96.7% success rate on real tasks, on par with GPT-6 Astra under the same testing framework, higher than MolmoAct2 VLA's 33.3%, and runs 3 times faster than GPT-6 Astra at 1/17 of its cost. On a single-arm Franka, ZAP scores 55.9% on RoboLab-120, 4.9 percentage points above the leading VLA/WAM baseline. These figures are all results reported by the authors.
Report timeline
- ZAP: Pretraining handles understanding, simulation training teaches robot motion
The authors propose Zero-real Agent Policy (ZAP), which uses a pretrained 27B-parameter vision-language model with a simple tool interface to control a robot, trained via supervised fine-tuning and GRPO on simple simulated tasks, without using any real robot data during training.
arXiv cs.RO ·