RoboScript Robot Learning Hits 83.4% From One Demo Per Task

A new system called RoboScript teaches robots each task from a single demonstration plus practice in a simulator, and hit 83.4% success across 50 robot tasks, beating the best trained robot policy by more than 20 points without retraining the AI model.

Key Takeaways

RoboScript reached 83.4% average success on all 50 tasks of the RoboTwin 2.0 robot benchmark, from one demonstration per task
That beats the strongest trained robot policy by more than 20 points, with no changes to the AI model’s weights
The method also worked on a physical two-armed robot
The researchers have released all the code

Researchers Heng Fang, Zehui Guo, Yuehan Zhang, Mingyang Du, Dingkang Liang and Xiang Bai describe the system in a paper posted on Nov. 9.

Practice Instead Of Retraining

Robots usually learn a skill by retraining their model, which ties the skill to the setting it was trained in. RoboScript keeps the AI model frozen. It watches one demonstration, rehearses the task in a simulator, learns from its mistakes, and saves what it learned as a task memory written into its instructions rather than its weights.

The researchers say this works because what the robot learns is about steps and the roles of objects, not how a scene looks, so the memory carries over to new scenes.

Also Read: Humanoid Robots Running Google’s Gemini Just Did Factory Work At Mercedes-Benz, And Got 96% Right

What The Results Show

The 83.4% covers both clean and randomized versions of the benchmark tasks. The team says results improved when the system reflected on its failures and got better with stronger AI models, and the memory could be adapted to a real dual-arm robot.

The results come from one benchmark, mostly in simulation, so they do not show that robots can now learn any real-world job from one example.