Neuro-Symbolic Imitation Learning: Discovering Symbolic Abstractions for Skill Learning

TL;DR

This work tackles the challenge of long-horizon robotic imitation by learning a relational symbolic abstraction and neural skills from task demonstrations. It jointly learns a predicate set $P$ and an operator set $\Sigma$ to enable abstract planning, and then trains neural skills $\Pi$ to refine abstract plans into executable actions. Abstract plans are generated with off-the-shelf planning and selected to align with demonstrations, after which neural skills execute the plan through subgoal sampling and goal-conditioned control. Across three simulated environments, the proposed neuro-symbolic framework demonstrates improved data efficiency, stronger generalization to unseen goals and object counts, and enhanced interpretability relative to baselines.

Abstract

Imitation learning is a popular method for teaching robots new behaviors. However, most existing methods focus on teaching short, isolated skills rather than long, multi-step tasks. To bridge this gap, imitation learning algorithms must not only learn individual skills but also an abstract understanding of how to sequence these skills to perform extended tasks effectively. This paper addresses this challenge by proposing a neuro-symbolic imitation learning framework. Using task demonstrations, the system first learns a symbolic representation that abstracts the low-level state-action space. The learned representation decomposes a task into easier subtasks and allows the system to leverage symbolic planning to generate abstract plans. Subsequently, the system utilizes this task decomposition to learn a set of neural skills capable of refining abstract plans into actionable robot commands. Experimental results in three simulated robotic environments demonstrate that, compared to baselines, our neuro-symbolic approach increases data efficiency, improves generalization capabilities, and facilitates interpretability.

Figures (4)

• Figure 1: (Top Left) The components of the neuro-symbolic policy. Predicates abstract the state-space, operators define an abstract transition model, and skills execute abstract plans. (Top Right) Illustration of the policy execution. First, the start and goal state are abstracted using the predicates. Following, an abstract plan is computed using the operators and planning algorithms. Lastly, the corresponding skill sequence is executed. (Middle) Overview of the learning pipeline. First, a set of candidate abstractions is generated based on the demonstrations. Subsequently, a subset of these candidates is selected using a novel objective function. Lastly, the learned symbolic representation is utilized to learn a set of skills with behavior cloning. (Bottom) The evaluation tasks. In each task, a panda robot has to manipulate objects placed on a table.
• Figure 2: Visualization of learned predicates. Predicates are visualized by overlaying images of states in which the predicate is true.
• Figure 3: Illustration of learned operators. The operators are shown in PDDL-Syntax.
• Figure 4: Comparison between our approach and baselines in three robotic environments. The x-axis denotes the different generalization scenarios; the y-axis denotes the success rate. Results are averaged over 10 random seeds, with bars representing the mean success rate and black lines indicating the standard deviation. Color shades denote the number of demonstrations used during training.