VQ-CNMP: Neuro-Symbolic Skill Learning for Bi-Level Planning

Abstract: This paper proposes a novel neural network model capable of discovering high-level skill representations from unlabeled demonstration data. We also propose a bi-level planning pipeline that utilizes our model using a gradient-based planning approach. While extracting high-level representations, our model also preserves the low-level information, which can be used for low-level action planning. In the experiments, we tested the skill discovery performance of our model under different conditions, tested whether Multi-Modal LLMs can be utilized to label the learned high-level skill representations, and finally tested the high-level and low-level planning performance of our pipeline.

• The paper presents a neural network architecture that learns high-level skill representations from unlabeled demonstrations using vector quantization.
• The paper formulates a bi-level planning pipeline that integrates high-level decision-making with LLM-based skill labeling and gradient-based low-level control.
• The experiments demonstrate up to 98% skill clustering accuracy in realistic environments, highlighting the method’s robustness and scalability.

Neuro-Symbolic Skill Learning for Bi-Level Planning in Robotics

The paper "VQ-CNMP: Neuro-Symbolic Skill Learning for Bi-Level Planning" introduces a novel approach to learning high-level skill representations from unlabeled demonstration data using a neural network model. The authors propose a bi-level planning pipeline leveraging these skill representations in a gradient-based planning framework. This bi-level approach is designed to effectively separate high-level decision-making from low-level perception and control, thereby enhancing robotic planning efficiency across various environments.

Key Contributions

The primary contributions of this paper include:

1. Skill Discovery Method: The paper presents a neural network architecture capable of learning high-level skill representations while retaining low-level action information. This approach aims to cluster demonstrations into discrete skills in an unsupervised manner, facilitating long-horizon planning tasks.
2. Bi-Level Planning Pipeline: The authors formulate a bi-level planning method that utilizes both learned high-level skill representations and low-level detailed planning to address complex tasks. This pipeline encompasses skill discovery, labeling, and planning using a combination of expert input and Multi-Modal LLMs.

Methodology

Model Architecture

The model employs a vector-quantized autoencoder architecture to cluster high-level skills from demonstration datasets. It uses conditional neural movement primitives (CNMPs) and vector quantization to map different skill variations onto discrete vectors within a learned skill space. The model integrates the benefits of continuous motion trajectories and discrete skill recognition, optimizing both for high-level task abstraction and detailed action execution.

Planning Approach

The paper details a multi-step process involving clustering demonstrations, skill labeling using LLMs, high-level planning with an LLM-based agent, and low-level planning using a gradient-based method. The planning system is designed to execute detailed actions derived from high-level plans made possible by the robust abstraction capability of the model.

Experimental Insights

The experiments conducted in a kitchen environment showcase the efficacy of the proposed method in classifying and planning skills such as retrieving and interacting with objects. The model effectively clusters demonstrations with high accuracy even under unequal dataset conditions, indicating resilience and adaptability.

1. Skill Discovery Performance: The model's high accuracy in clustering skill demonstrations (up to 98% in some cases) highlights its effectiveness in unsupervised skill learning. The study emphasizes model consistency across different skill space sizes, providing insights into scalability.
2. Multi-Modal LLM Utilization: By leveraging LLMs for skill labeling, the paper explores automation in the bi-level planning pipeline. The results indicate potential but also suggest that LLMs require further enhancement for reliable automated skill labeling.
3. Planning Performance: High-level planning demonstrated varied performance depending on prompt clarity and environment understanding. Notably, transparency about object locations significantly improved planning outcomes, underscoring the importance of precise task environments.

Implications and Future Directions

The work presents significant implications for robotics, particularly in integrating LLMs for task reasoning and automated skill labeling. The neuro-symbolic approach bridges the gap between abstract task definitions and concrete robotic actions, illustrating a promising direction for future AI developments.

However, future research should focus on refining the interaction between LLMs and skill representations, exploring more reliable state abstractions, and expanding to larger datasets and more complex task environments. By enhancing LLM capabilities and refining the bi-level planning framework, broader and more adaptable robotic applications can be realized.