[
Ambi Robotics Inc., a developer of material handling systems using artificial intelligence, today launched PRIME-1, its robotic foundation model for commercial warehouse operations. The Berkeley, Calif.-based company said the model can improve performance, accelerate product development, and increase the reliability and maintainability of its sorting robots.
Recent advances in generative AI and large language models have opened new doors for robotics developers. An increasing number of companies has released foundational models like PRIME-1 for training robots.
Last year, NVIDIA announced Project GR00T, a research initiative developing general-purpose foundation models, tools, and technologies for accelerating humanoid robot development. Like GR00T, many foundation models released in the past year focused on training humanoid robots.
Ambi is taking a different approach. PRIME-1, which stands for Production-Ready Industrial Manipulation Expert, provides a unified transformer backbone that can be fine-tuned for a variety of robot operations including 3D perception, package picking, and quality control.
“With PRIME-1, we’re addressing the most pressing challenge in warehouse robotics: the need for adaptable, scalable solutions that evolve with operational demands and amplify return-on-investment,” said Jeff Mahler, co-founder and chief technology officer of Ambi Robotics.
“PRIME-1 allows our customers to leverage collective learning from our entire production fleet, empowering them to stay ahead in the rapidly evolving logistics landscape with increasing demand,” he added. “Our customers now have the ability to respond faster to market dynamics and future-proof their operations in an industry where speed and precision are paramount.”
Real-world data provides better training opportunities
Last month, Mahler told Automated Warehouse that Ambi Robotics continually improves the AI that runs its robots. At the time, he said the company typically rolls out upgraded AI models, equipped with more recent data, about once a month. Now, it is taking this a step further with its foundation model.
Ambi pre-trained PRIME-1 with self-supervised deep learning on over 20 million high-quality images from individual pick, place, and pack events. These events span 150,000 operating hours across the company’s AI-powered sortation fleet in U.S. warehouses. The company said this training dataset represents about 1% of the data it has collected to date.
Ambi said the result is that PRIME-1’s training leverages a reliable depth of real warehouse data from identical systems deployed at scale. The breadth and specificity of the data ensure that PRIME-1 is optimized for precision and efficiency in real-world logistics operations, it asserted.
“The scale of training on real-world data collected from the AmbiSort A-Series systems has enabled us to reach high levels of reliability with PRIME-1,” said Vishal Satish, foundation model lead at Ambi Robotics. “The use of PRIME-1 will allow us to rapidly develop and deploy new robotic solutions for a variety of tasks, while also improving the performance of our existing robotic systems.”
PRIME-1 prepares Ambi for deployment at scale
PRIME-1, trained on over 1 trillion tokens, has learned generalizable features for 3D reasoning, which can be applied to a range of challenging 3D tasks, including depth estimation and robotic picking.
“Emerging AI research shows that generative pre-trained models can outperform previous architectures,” said Ken Goldberg, co-founder and chief scientist of Ambi Robotics. “ The engineering team at Ambi Robotics used four years of proprietary warehouse data to train a state-of-the-art generative model for 3D warehouse operations. Their experiments with real production systems confirm that PRIME-1 significantly outperforms their previous systems.”
Testing in production has revealed clear scaling laws, showing that both pre-training quality and performance on downstream tasks improve as the amount of data used for pre-training increases. This suggests that pre-training on a large volume of unlabeled data can significantly enhance performance.
Pre-training with unlabeled data surpassed the results achievable with labeled data alone, noted Ambi Robotics. The company said it has yet to see performance saturation.