A new AI framework teaches household robots to learn from their own screwups, no human grading required.
Researchers built MobiAgent, a two-loop system for robots that both walk and manipulate objects. The inner loop uses vision-language models to plan step by step and catch mistakes mid-task, breaking jobs into reusable skills handled by specialized control modules. The outer loop reviews the robot's own recorded attempts, sorts them into skill categories, and retrains the skill library automatically, no human annotation needed. In tests on the RoboCasa and BEHAVIOR-1K simulated benchmarks plus real hardware, MobiAgent beat a baseline model called pi-0.5-TA by 22.5 percentage points on BEHAVIOR-1K, and its self-training loop lifted RoboCasa success from 7.5% to 27.5% and Astribot S1 success from 32.5% to 57.5%.
Most robot AI handles single, short tasks fine but falls apart on multi-step chores, like clearing a table and then loading a dishwasher, because small errors compound across stages. MobiAgent's fix is giving the robot a feedback loop that mines its own failed attempts for training data, instead of waiting on engineers to hand-label more examples. That is the same self-supervised flywheel that pushed large language models forward, now pointed at the messier problem of machines that move and grab things at once.
The percentage gains are real, but a robot that succeeds just over half the time on Astribot S1 is still failing nearly as often as it wins - progress, not a finished product.