Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Zero-real Agent Policy (ZAP) is a 27B-parameter vision-language model post-trained to control robots through a minimal tool harness. The harness sets end-effector poses and gripper openings, with no grasp macros or motion primitives, and the same harness runs in simulation and on hardware. The model writes a plan, names objects to locate, then issues perception and movement tool calls, seeing fresh camera images and robot state on every turn.
Training uses only simulated, deliberately simple pick-and-place and stacking tasks. The model is fine-tuned on its own successful rollouts and then optimized with GRPO. There is no frontier teacher for distillation and no real-world robot data (teleoperation, demonstrations or real rollouts). The project reports zero-shot transfer to a real bimanual robot on tasks with unseen objects and goals, including placing a test tube in a rack, a mid-air bimanual handoff, drawing an X on a tic-tac-toe board and making a checkmating chess move.
On real-world tasks selected from RoboLab-120, the authors report that ZAP matches GPT-6 Astra running in the same harness. They report 146 s per task versus 397 s, and $0.29 versus $4.87 per task. Turn counts are similar (21.9 vs 19.4). These are the authors' own figures. Based on the material provided, only a project page and paper are available; no code, weights or product access were found, and the vendor is not identified in the provided text.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
27B vision-language-action robot policy trained entirely in simulation with RL on 40 simple tasks, yet performs unseen tasks on real robots, matching GPT-6 Astra's success rate faster and cheaper. Only a paper released so far.
Closest neighbours by meaning, not just by category.
Library
Embodied harness from NUS Show Lab that lets vision-language models control robots via discrete semantic action units.
Research
Research framework letting a frozen VLM agent learn robot tasks in context from videos, images and feedback, no fine-tuning.
Model · Perceptron
Perceptron's vision-language model for physical AI, covering perception, physical reasoning and robot action prediction.
Model
Action-conditioned robotics world model fine-tuned from Cosmos-Predict2.5-2B using 2D skeleton conditioning across embodiments.
Model · Sunday Robotics
Robot control model from Sunday Robotics for household tasks such as laundry folding, reportedly running offline.
Model · Unitree
Unitree's 6B-parameter vision-language-action foundation model for humanoid robots, covering 64 tabletop and whole-body tasks.
