Robots learning from demonstrations need diverse data to handle unexpected situations. Current methods either rely on humans for safe but limited data or let robots explore on their own, which can be unsafe and inefficient. Data Inoculation combines both: a human teleoperates the robot while a learned policy adds small, controlled deviations, mimicking potential mistakes. The human then corrects these deviations, providing recovery demonstrations. This creates a dataset that includes both errors and corrections, making the final policy more robust. On three real-world tasks, it improved success rates for multiple learning methods. This matters because it offers a practical way to collect higher-quality data without sacrificing safety or requiring many failures.
Robots can learn better when humans correct deliberately introduced deviations
Across three real-world tasks, the authors report that adding learned perturbations during human-controlled demonstrations produced training data that improved or matched downstream robot success rates.
Top University
Qiyang Li · Zhiyuan Zhou · Charles Xu · Sergey Levine
UC Berkeley
Research Digest··2 min read
In one line
Data Inoculation uses a shared-control framework to perturb demonstrations, improving downstream policy robustness.
Before you read this
An idea this paper builds onRead it as
Li and colleagues investigated whether robot demonstrations become more useful when humans must continuously correct a learned controller’s deviations.
The authors developed Data Inoculation, a data-collection strategy in which a human teleoperates a robot while a learned controller adds perturbations to the human’s commands.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.