Knowledge-Guided Manipulation Using Multi-Task Reinforcement Learning

1. 1Augment egocentric vision with an online 3D scene graph
2. 2Update spatial, containment, and other relations using a dynamic-relation mechanism
3. 3Use multi-task model-based policy optimization to learn policies for various manipulation tasks
4. 4Integrate knowledge graph with the policy optimization framework to guide manipulation decisions

Robotics engineers and AI researchers on a team can benefit from this framework as it enables more efficient and accurate manipulation tasks, while also providing a unified approach to perception, knowledge, and policy

💡 Unifying perception, knowledge, and policy using a knowledge graph and multi-task reinforcement learning can improve robotic manipulation in partially observable settings

💡 Knowledge-Guided Manipulation uses multi-task RL for robotic manipulation