Deep latent-space sequential skill chaining from incomplete demonstrations

Abstract

Imitation learning is a methodology, which trains an agent using demonstrations from skilled experts without external rewards. However, for a complex task with a long horizon, it is challenging to obtain data that exactly match the desired task. In general, humans can easily assign a sequence of simple tasks for performing complex tasks. If a person gives an agent an order of simple tasks to carry out a complex task, we can find a skill sequence efficiently by learning the corresponding skills. However, independently trained low-level skills (simple tasks) are incompatible, so they cannot be performed in sequence without additional refinement. In this context, we propose a method to create a skill chain by connecting independently learned skills. For connecting two consecutive low-level policies, we need to find a new policy defined as a bridge skill. To train a bridge skill, a well-designed reward function is required, but in real world, only sparse rewards can be given according to the success of the overall task. To complement this issue, we introduce a novel latent-distance reward function from fragmented demonstrations. Also, we use binary classifiers to determine whether the current state is capable of performing the skill that follows. As a result, the skill chain formed from incomplete demonstrations can successfully perform complex tasks which require performing multiple skills in a sequence. In the experiment, we solve manipulation tasks with RGBD images as input in the Baxter simulator implemented using MuJoCo. We verify that skill chains can be successfully trained from incomplete data while confirming that the agent can be trained much more efficiently and stably through the proposed latent-distance rewards. Also, we perform block stacking using a real Baxter robot in the simple set-up environment.