Learning deep policies for physics-based robotic manipulation in cluttered real-world environments

This thesis presents a series of planners and learning algorithms for real-world manipulation in clutter. The focus is on interleaving real-world execution with look-ahead planning in simulation as an effective way to address the uncertainty arising from complex physics interactions and occlusions.

We introduce VisualRHP, a receding horizon planner in the image space guided by a learned heuristic. VisualRHP generates, in closed-loop, prehensile and non-prehensile manipulation actions to manipulate a desired object in clutter while avoiding dropping obstacle objects off the edge of the manipulation surface. To acquire the heuristic of VisualRHP, we develop deep imitation learning and deep reinforcement learning algorithms specifically tailored for environments with complex dynamics and requiring long-term sequential decision making. The learned heuristic ensures generalization over different environment settings and transferability of manipulation skills to different desired objects in the real world.

In the second part of this thesis, we integrate VisualRHP with a learnable object pose estimator to guide the search for an occluded desired object. This hybrid approach harnesses neural networks with convolution and recurrent structures to capture relevant information from the history of partial observation to guide VisualRHP future actions.

We run an ablation study over the different component of VisualRHP and compare it with model-free and model-based alternatives. We run experiments in different simulation environments and real-world settings. The results show that by trading a small computation time for heuristic-guided look-ahead planning, VisualRHP delivers a more robust and efficient behaviour compared to alternative state-of-the-art approaches while still operating in near real-time.