We've been at it for a couple of months now @nomagicAI and want to share our progress. Our thesis for physical AI is covered in detail in our blog and the articles below.
We need to be humble on the path to useful, broadly-capable robots. AI research has demonstrated what scaling can achieve, and classical robotics has built systems that clear production reliability. Each side holds a piece, and progress requires the best of both worlds, not either side pushing forward alone. Our key hypotheses:
Mastery before generality: The field is racing to build the most general robot brain. We're betting the harder part is mastery, and it needs focus now. The last stretch from impressive demo to 99.9%+ autonomy is won in production, not pretraining. Robotics' "ChatGPT moment" won't come from a single breakthrough model; it arrives via deployment.
Grounding in deployment: Classical deployment data is robotics' internet-scale data, generated right now by robots doing real work. Deployment grounds our benchmarks too: we measure what we actually care about, not proxies.
The classical stack as harness: Every general physical AI model alone sits below production-grade reliability. Harnesses are hard for physical AI, but the classical stack provides the key functions and keeps the system above the autonomy bar while the model improves. This enables cheap patching via a few demonstrations, classical-stack trajectories, or targeted simulation, and lets RL focus on exactly the problems that matter. The harness doesn't just protect the model, it bootstraps the whole process.
Warehouse before living room: Physical AI will mature in warehouses and factories long before the living room. The conditions are decidedly better for data and evaluation. Sim-to-real & teleop alone buy reasonable performance at acceptable cost. But in a warehouse, reasonable is useless: no one wants your system if a human has to step onto the floor even once an hour. We're already at millions of real picks/month across our fleet and growing exponentially, over 2M/month with Zalando alone (featured in their Q1 earnings: https://corporate.zalando.com/en/financials/zalando-q1-2026-results).
Early evidence: We've deployed what we believe is the world's first item-manipulation VLA in production, with paying customers. Not a pilot, regular deployment, and it will only get better. For Brack.Alltron this means robot-caused human interventions roughly halved, enabling autonomous shifts through nights & weekends. The same flywheel will drive transfer across tools, starting with our shoebox gripper. This path has already carried our perception systems past operating points generalist models do not reach in production. More on that soon.
What comes next: We're growing the Zurich and Warsaw labs, hiring world-class researchers & engineers 👇 (more in our blog and the articles below)
#PhysicalAI #Robotics #VLA #EmbodiedAI #FoundationModels #WarehouseAutomation