FOCUS: object-centric world models for robotic manipulation.

Area of Science:

• Robotics
• Artificial Intelligence
• Cognitive Science

Background:

• Current reinforcement learning (RL) world models often use global latent vectors, lacking object-specific structure.
• This limits the ability to understand and predict object interactions, crucial for complex tasks.

Purpose of the Study:

• To develop an object-centric world model for reinforcement learning agents.
• To design an object-centric exploration mechanism to enhance interaction discovery.
• To improve robotic manipulation skills and learning efficiency.

Main Methods:

• Proposed FOCUS, a model-based agent learning an object-centric world model.
• Implemented an object-centric exploration strategy encouraging interaction with scene objects.
• Benchmarked performance in robotic manipulation tasks.

Main Results:

• FOCUS achieved more accurate object predictions and efficient learning.
• The object-centric exploration strategy promoted diverse interactions (reaching, moving, rotating).
• Demonstrated fast adaptation to sparse reward RL tasks and real-world applicability with a Franka Emika robot arm.

Conclusions:

• Object-centric world models significantly improve RL agent performance in robotic manipulation.
• Focused exploration on objects accelerates learning and task adaptation.
• FOCUS offers a promising direction for real-world robotic applications.