Ambi Robotics launches PRIME-1 trained on 20M warehouse images and 150,000 robot hours

Ambi Robotics launched PRIME-1 on 7 January 2025, introducing a transformer-based foundation model trained on 20 million images from more than 150,000 hours of warehouse robot operation. The company said the model was already running in commercial US warehouses, supporting three-dimensional perception, package picking and quality control.

PRIME-1 uses a shared transformer backbone that can be fine-tuned for different robot functions. Its self-supervised training draws on images of individual pick, place and pack events collected by AmbiSort A-Series systems. Ambi said the training dataset represented about 1% of all data it had collected and contained more than one trillion tokens.

The model learns features for tasks such as depth estimation and robotic picking from four years of production data. Ambi reported that both pre-training and downstream-task performance improved as more data was added, with no saturation observed in its experiments. The company intends to use the model to improve existing sorting systems and accelerate development of further warehouse applications.