AutoRoboHarness

AutoRoboHarness is a team of robot agents that improves itself. Robots work tasks in a simulated retail store by reading a shared, hierarchical brain, a knowledge graph of guardrails, rules, protocols and procedures. After every shift, a Nemotron 3.5 Super VL supervisor reviews the robots' trajectories against real human footage from the RetailAction dataset and against sibling runs. It finds failures and proposes brain changes as graph operations, citing published practice through You.com. A separate DeepSeek V4 Pro judge drops any claim the evidence doesn't support. A gate then re-runs 10 held-out scenarios, which the supervisor never sees, with the old brain and the new one, and it keeps a change only if the score rises. In our verified run, one weight-based grasp rule lifted held-out completion from 6/10 to 9/10 and cut refused actions per task from 0.5 to 0.2. A fresh comparison reproduced the 6/10 against 9/10 result. Users launch up to 4 Nemotron 3.5 Lightning robots, watch them in 3D or through head cameras, and see every brain version's held-out score on a history chart.