Store-Picker-Bot: From Natural Language Orders to Real Robot Picks
Our team has hit some key milestones in building Store-Picker-Bot for the Nebius x NVIDIA Global AI Hackathon 2026. Here's where the project stands today.
Store-Picker-Bot is an AI-powered robot that picks grocery orders autonomously, the way quick-commerce dark stores work. A customer places an order in plain language, NVIDIA Nemotron plans the job, and a mobile robot with an SO-100 arm, running NVIDIA GR00T on a Jetson AGX Orin, picks the items into its basket.
Web app: completed The QuickPick web app is finished and working:
- Customer ordering page: customers can pick products or type an order, pay (demo), and follow their order live.
- Owner dashboard: shows the live robot map, Nemotron's step-by-step log, orders, stock, KPIs, and has an emergency STOP.
- Demo mode: the full flow runs with a simulated robot, so the app can be shown and tested without hardware.
Nemotron: integrated NVIDIA Nemotron is fully integrated through Nebius Token Factory:
- Nemotron Nano reads customer orders and turns them into structured items.
- Nemotron Super works as the planning agent. It checks stock, substitutes items, plans the shelf route and sends step-by-step commands to the robot.
- Nemotron Ultra writes the end-of-shift summary for the store owner.
Every robot command from Nemotron goes through safety checks (no over-picking, no picking at the wrong shelf). If there is no API connection, an offline planner takes over automatically.
GR00T arm policy: showing good results We fine-tuned GR00T N1.7 on real teleoperated demonstrations recorded on our own robot, with no simulation. Today the policy started showing good results on the arm, which is a major milestone for the project.
Testing All test suites pass, and a full order (order, plan, robot picking, completed) works end to end in demo mode.
Next steps
- The only part left is connecting the web app to the real robot. Now that the GR00T policy is performing well, this is our next step.
- After that, we'll test the complete flow on the robot, from a customer's order to a packed basket.
Thanks to everyone following our progress. We're excited to show the full system working on the real robot soon!