Masked IRL: LLM-Guided Reward Disambiguation from Demonstrations and Language

1. 1Collect demonstrations of a task
2. 2Use LLMs to identify relevant state details
3. 3Disambiguate reward functions using language and demonstrations
4. 4Fine-tune the reward model for improved generalization

ML researchers and roboticists can benefit from this approach to improve the accuracy of reward models and enable more effective human-robot collaboration

💡 LLMs can help disambiguate reward functions by identifying relevant state details and specifying what matters for a task

🤖 Masked IRL uses LLMs to improve robot learning from demos & language!