Researchers from MIT, NYU, and UC Berkeley have developed a system enabling rapid robot task learning with minimal human effort. As described in an MIT press release, the technique allows non-technical users to easily train robots for new duties via intuitive feedback. This marks a major step toward adaptable general purpose robots that can smoothly handle novel real-world situations.

Consider household robots pre-trained at factories for fixed abilities confronting unfamiliar settings and objects. If a robot fails to grasp a new mug, opaque programming means users can’t deduce why. "The key missing component is the ability for the robot to show the reasons behind failure so the user can provide feedback," says MIT’s Andy Peng.

Their system bridges this gap. When the robot falters, it generates counterfactual explanations describing tweaks for success, e.g. mug color changes. The human reviews these suggestions and critiques the failure reasons. This feedback becomes training data to rapidly fine-tune the robot's model.

In simulations, the technique proved superior for quickly adapting robots versus other methods, with less human time required. It helps robots rapidly acclimate to new environments without needing technical expertise. This could eventually enable capable general purpose robots to smoothly perform everyday tasks for seniors or disabled users in varied circumstances.

Looking ahead, the researchers hope to trial the system on physical robots. They also aim to leverage generative machine learning to speed up training data generation. "We want robots that perform in a semantically meaningful human-like way, not just optimizing pixel-level properties,” says Peng.

By enabling rapid human-centered learning, the work brings closer adaptable yet trustworthy robots that understand failures and can smoothly take on new duties. The research will be presented at the International Conference on Machine Learning.