How Weights & Biases helps Siemens empower warehouse robots with advanced AI automation
Siemens has been one of the world’s foremost multinational technology conglomerates since its founding in Germany in 1847. With over 300,000 global employees today, the company continues to transform industrial automation, distributed energy resources, rail transport, and health technology worldwide.
For Product Portfolio Manager Christopher Schuette and his Digital Industries team, staying at the forefront of digital technologies is what Siemens prides itself on. It is no surprise that their forays into AI and robotics as integral parts of industrial automation have been successful trailblazing efforts.
“We are in an era of digitization, where AI is impacting and transforming industries worldwide,” said Schuette. “Amazing tools like Weights & Biases empower our exceptional team of AI and robotics researchers, as well as ML engineers, to tackle highly complex challenges and deliver innovative solutions for our users with efficient time-to-market.”
“This enables our team to strengthen our data flywheel and brings us closer to our vision of becoming the world’s leading technology partner for user-friendly advanced robotics applications in industrial automation. We are helping automation and robotics engineers worldwide develop new generations of intelligent robot solutions in a scalable way on one of the world’s leading platforms for industrial automation, enabling them to leverage these groundbreaking technologies to boost innovation and growth in their business.”
Schuette’s Factory Automation team – one of the leading business units within the broader Digital Industries team – developed SIMATIC Robot Pick AI, a Foundation Model-based software for Robotic Piece Picking solutions. This software is transforming standard robots into intelligent robots, allowing them to make adaptive and reliable decisions on how to pick items in the high variety of Stock Keeping Units (SKUs) with different shapes and sizes without prior information. This has made for example warehouse operations much more efficient by reducing dependencies on today’s repetitive manual-only tasks in a time of rising labor constraints within this domain.
Weights & Biases has been a critical part of the model development pipeline at Siemens, and played a pivotal role in the continued development and improvement of SIMATIC Robot Pick AI. Learn more about Siemens’ cutting edge robotics technology, the challenges they faced, and how Weights & Biases helped them overcome those obstacles.
SIMATIC Robot Pick AI: Delivering operational efficiency in complex warehouse environments
Siemens identified a critical challenge in intra-logistics operations: maintaining operational efficiency amid growing labor scarcity. Unlike synchronized production lines, intra-logistics represent highly complex environments where items constantly change in size, format, and packaging—all while consumer behaviors and e-commerce demand continue to evolve. With over 99% of order fulfillment requiring human hands for picking and packing tasks, particularly during high-volume periods like Black Friday, the industry faces a significant dilemma as the labor pool shrinks in industrialized countries.
Traditional automation has been severely limited by rule-based approaches, failing when object variability is high. While robot arms can operate fixed programs in physical environments, they’ve lacked the ability to truly understand and adapt to their surroundings. SIMATIC Robot Pick AI was developed to solve this challenge.
SIMATIC Robot Pick AI inputs depth images, and color images, enabling robotic arms of any kind to have the intelligence and capability to find reliable pick poses on arbitrary novel items. This eliminated the need for item-specific training, delivering autonomous behavior that went beyond conventional rules-based automation.
“We want to make market-ready, value-driven, ready-to-use industrial applications such as SIMATIC Robot Pick AI to be used by non-AI experts,” said Schuette. “Imagine individuals who do not fully understand the intricacies behind AI; we want to enable these robot solution providers in their warehouse market to enrich their portfolio with adaptive and intelligent robot solutions, based on AI systems developed using W&B.”
How Weights & Biases improved Siemens’ ML workflow
“Before using W&B Models, there were several pain points at different parts of our ML development workflow,” said ML engineer Kyle Coelho. “One was definitely when initiating runs, we wanted to run streamlined training sweeps to really help us iterate much faster. Secondly, we didn’t have an easy way to track lineage. Lastly, we were slower in making decisions because everything was local on our on-prem servers so we basically had to run custom scripts to generate visualization and manually iterate through them.”
The biggest issue pre-W&B was that everything was done in a largely manual fashion, from initiating training runs to monitoring metrics and validation losses, all the way to benchmarking and choosing candidate models. Everything was done on local on-prem servers, with no single platform or single go-to source to track everything, which naturally created a lot of inefficiencies. Weights & Biases solved many of them.
The first thing W&B Models helped Coelho solve was when running hyperparameter sweeps. Being able to iterate quickly and use W&B visualizations allowed the team to narrow in much faster on the right hyperparameters to use for each model. As an added bonus, once they narrowed in on that, they could also have a better ballpark of what future training runs should look like.
Coelho and the team also took advantage of W&B’s full lineage tracking on all models to track everything, from the configuration that was used to the model artifacts that were generated. Being able to identify for any given model release which run and config was used to reach that point helped the team establish solid benchmarks. And with Tables, the team no longer had to generate custom scripts just to analyze plots, but could instead customize whatever visualizations they wanted within W&B, helping them narrow down on candidate models much faster.
W&B now plays an essential role within the Siemens model development pipeline. The workflow starts with data sources pooled into Siemens’ servers. The team then uses on-prem NVIDIA clusters to support the labeling processes and once that is complete, W&B takes over throughout the training and experimentation process, right up until releasing a final model candidate to be used downstream in Registry.
“Once we have a bunch of trained models, we log multiple custom metrics and use Tables for visualizations and benchmarking each of the models that goes into the product,” explained Coelho. “Then we have a manual review of all these metrics, and we use Registry to tag v1, v2, v3, whichever release for the right model to be promoted into the subsequent testing and integration phases prior to deployment, and eventually packed into a Pick AI release.”
In addition to Coelho and his ML engineering team, the Siemens Digital Industries team also benefits from W&B Reports, using it to communicate results broadly across the team and collaborate on analysis.
“I consume Kyle’s Reports when we have a particular problem we’re trying to solve, for example a systematic issue with our neural networks,” said Tech Lead and Senior Key Expert Ines Ugalde. “Everything is laid out in there – this is what we tried, this is what worked, what didn’t work well. And it’s all recorded, it’s clickable, I can go in and see the runs and configs. It really helps us debug and solve problems much faster. The information is neatly organized, it gives us incredible speed and efficiency and that makes our delivery process smooth and intuitive.”
“Weights & Biases has given us a lot of peace of mind,” summed up Coelho. “We’re able to minimize the smallest possible likelihood of things slipping through the cracks. It’s not only been a huge productivity boost, but it gives me peace of mind knowing exactly what’s going on throughout our model development pipelines.”