Problem: The Navigation Among Movable Obstacle (NAMO) problem, where agents can move objects in its environments to reach otherwise unreachable positions.

Methodology: We reframed the NAMO problem into a reinforcement learning problem, and used a multimodal deep neural network for function approximation.

Technical Detail: The network was implemented and trained with PyTorch, and NVIDIA Isaac Gym was used for physics simulation. Custom environments were built to generate random NAMO scenes.

Results: The final policy is able to understand the relationships between the object, the room configuration, and its goal. It generalises to unseen rooms and object positions, and we tested the network on a quadrupedal robot.