11/19/2026 | Press release | Distributed by Public on 11/19/2026 10:24
Industrial battery sorting has historically been a manual process. An operator picks up a battery, reads the label, and makes a call. That works at low volume. It doesn't work when you're building a domestic battery supply chain that processes tens of thousands of metric tons of material a year. So, we built a line that sorts automatically, using computer vision and machine learning, and it's now sorting real feedstock for our recycling operations.
Why battery sorting matters
Much of what we sort feeds our rotary calciner, which takes live battery cells, consumer electronics batteries, and EV modules as raw feedstock. It is designed to process over 40,000 metric tons a year, or roughly 15 to 20 GWh of battery capacity.
A thermal process at that scale runs best when it knows what's going in. Better upstream sorting means less variability in the process, higher recovery rates, and less wear on the equipment.
So, when our engineering team looked at how to build a better sorting process that was more accurate, more efficient, and could scale, we knew we needed to automate our sorting line. Our engineering team built a system that can detect variability throughout a multi-stage pipeline and make a judgement call by applying machine learning and computer vision.
The automated sorting line is what makes that possible at industrial throughput and is already processing real feedstock for our recycling operations, improving our recycled metals quality and smoothing our recycling processes.
The challenge: from a pile to a line
Before any sensing or classification can happen, the line has to solve a messier problem. What arrives at Redwood from collection partners across the country is not a neat queue of batteries. It's a three-dimensional pile in drums, boxes, and bins, with batteries of every form factor and packed in insulating media. Humans don't sort accurately from a messy pile, and neither does a model.
Getting the batteries from pile to line is one of the less glamorous and more important pieces of engineering in the system. The infeed works the pile down in stages, separating out the insulation and narrowing the flow until what comes out the far end is a single-file line of batteries, moving through the system one at a time. Everything downstream depends on that line being clean.
The solution: recognizing each battery
Each battery passes through a sensor tunnel. Cameras capture high-resolution images of every side, including the bottom face, and resolve chemistry markings and label text. A laser profilometer measures the battery's geometry in three dimensions, and a checkweigher records its weight.
All sensor data streams asynchronously to an on-premise GPU cluster, where the inference pipeline classifies many batteries in parallel. Our models, trained on Redwood's own battery data, combined print text, shape, and weight to predict each battery's chemistry, and write that classification back to the controls system in real time. The models have also been optimized to return the classification by the time the batteries reach the outfeed of the line, even while processing at several metric tons per hour.
An earlier generation of the system used X-ray. Over time, we found that most of what we needed to identify chemistry was visible on the outside of the battery. Moving to standard cameras made the system cheaper, safer to operate, and much easier to scale.
Precise battery sorting
Routing each battery to the right destination is its own engineering problem. Once a classification lands in the controls system, the line sends each battery to its bin in real time. Different stations use different actuation, including mechanical diverters, pneumatics, and tightly coordinated servo motion control, each selected and tuned to keep up with production rate. A guide-by-light strip runs the length of each battery and shifts color by chemistry, giving operators a live read on what the line decided and the chance to pull anything that looks off before it moves on. Position tracking combines profilometer measurements and belt encoder data, which keeps the light locked to each battery even as belt speed varies.