Knowledge- and ambiguity-aware robot learning from corrective and evaluative feedback.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Interactive imitation learning (IIL) requires flexibility for non-expert users and must account for human error.
• Existing IIL methods often assume perfect teachers (oracles) and lack adaptability to diverse human factors.

Purpose of the Study:

• To propose an IIL method that enhances human-robot interaction for non-expert and imperfect teachers.
• To improve robot adaptability by incorporating uncertainty estimation and flexible feedback mechanisms.

Main Methods:

• Incorporated epistemic and aleatoric uncertainty estimation for the robot to identify knowledge gaps and demonstration ambiguity.
• Enabled teachers to provide corrective demonstrations, evaluative reinforcements, and implicit positive feedback.
• Developed an IIL framework adaptable to diverse teacher interaction preferences and error patterns.

Main Results:

• Demonstrated improved learning convergence compared to other methods, especially with ambiguous teachers.
• Experimental results showed enhanced data efficiency in the learning process.
• User studies indicated an improved overall teaching experience.

Conclusions:

• The proposed IIL method effectively addresses challenges posed by non-expert and imperfect teachers in human-robot interaction.
• Uncertainty estimation and flexible feedback mechanisms significantly improve robot learning and user experience.
• This work advances the deployment of adaptable robots for broader non-expert user adoption.