Sereact

Few processes in e-commerce are as hard to automate as returns. Every item comes back different, which is why returns processing has long been considered out of reach for robotics. Otto Group and Sereact, along with key partners like NVIDIA, are changing that. The first 10 Sereact dual-arm robots will be deployed at Otto Group’s logistics center in Poland to grasp, identify, assess and sort returns, including items they’ve never seen before. 15 more are planned in phase two, bringing the program to 25. This is what the future of physical AI looks like. Full announcement in comments.

No two returns look the same. That’s exactly why we built Cortex. Dual-arm picking, live at Zalando x CEVA Logistics.

Returns are among the hardest warehouse processes to automate. Every item arrives differently: refolded or crumpled, bagged or loose, tagged or not. No two items are alike. That means the robot has to keep learning. In the past, improving the policy meant weeks of collecting data, retraining and retesting. Not anymore. We’re teaming up with NVIDIA, Nebius, CEVA Logistics and Zalando to deploy a closed training loop in live returns operations. Now, improvements can reach the floor in days, not weeks. Watch how Cortex handles Zalando’s returns at a CEVA Logistics fulfillment center in Germany.

When we started Sereact, people told us returns were impossible to automate. Every item arrives differently: refolded or crumpled, bagged or loose, tagged or not. No two are alike, and that’s exactly why it’s one of the best tests of physical AI. This is Cortex handling Zalando’s returns at a CEVA Logistics' fulfillment center in Germany. It’s live and in production, every shift. Proud of the team that made this real.

We're bringing three stations to MODEX next week. A single-arm picking station that runs across thousands of SKUs. Fast, consistent, no breaks. A dual-arm returns station. Sorting, inspecting, re-slotting. The kind of work most warehouses still staff with temps because it burns people out within weeks. And a Lens station. AI based vision analysis you can interact with at the booth. All three running live at Booth B11311. If you're there, come watch them work.

The best place to train a robot is on the job. This week I joined the Physical AI panel at Bits & Pretzels with Lukas Köstler (SE3 Labs) and Milan Allan (alqem), moderated by Philipp Gerbert (TUM Venture Labs). We talked about what physical AI can already do, what is still hard, and where it goes next. My answer to the hard part: post-deployment continual learning. Most robots are trained in one world and deployed into another. Simulation is fast but never quite real. The lab is real but never quite the job. That gap is where robots fail. Keep learning after deployment, and that gap closes. Training and deployment distributions are aligned. In plain words: the robot learns in exactly the environment where it has to be good. Real lighting, real packaging, new products every week, and all the edge cases nobody would think to put in a test set. And when the job changes, the model changes with it. That's learning on the job. It's why deployment isn't the finish line for us at Sereact. It's where the most valuable learning starts. More than 90% of the episodes Cortex 2.0 was trained on came from real customer operations. Thanks Philipp for moderating, Lukas and Milan for a great discussion, and the Bits & Pretzels team for having us. #PhysicalAI #Robotics #bits26