Continuous Relaxation of Symbolic Planner for One-Shot Imitation Learning

De-An Huang, Danfei Xu, Yuke Zhu, Animesh Garg, Silvio Savarese, Li Fei-Fei, Juan Carlos Niebles

Abstract

One-shot imitation learning—the ability for a robot to acquire a new skill from a single demonstration—remains a significant challenge in robotics. Current end-to-end deep learning methods often require extensive meta-training data and struggle with long-horizon logic. Conversely, classical symbolic planners are robust in logic but fail when faced with the probabilistic and often noisy outputs of neural perception systems. This paper introduces a framework that bridges this gap by relaxing discrete symbolic planning into a continuous space. Our approach utilizes modular Symbol Grounding Networks (SGN) to map continuous states to probabilistic predicates and a Continuous Planner (CP) that operates directly on these distributions. Evaluated on complex block-stacking and sorting tasks, our method demonstrates superior data efficiency and robustness compared to state-of-the-art baselines.

Introduction

The "holy grail" of robotics is one-shot imitation learning, where a robot performs a complex task after seeing it done just once. However, existing "black-box" neural network policies typically conflate perception and control, requiring thousands of training examples to generalize effectively. While symbolic planners (e.g., PDDL) excel at long-horizon reasoning, they are notoriously brittle; they require discrete, error-free inputs that neural networks cannot reliably provide. Even minor uncertainty in a block's state can cause a traditional symbolic planner to fail.

Comparison of existing one-shot approaches, traditional symbolic planners, and the proposed continuous planner framework.

To address this, we propose a system that maintains the logical structure of symbolic planning while accommodating the "fuzzy" nature of deep learning outputs.

Methodology

Our framework consists of two primary components: modular Symbol Grounding Networks and a Continuous Planner.

Symbol Grounding Networks (SGN)

Instead of a monolithic policy, we employ modular neural networks to ground continuous environmental states (such as object poses) into symbolic predicates like On(A, B) or Clear(A). These modules are shared across different objects and tasks, significantly enhancing data efficiency.

The modular architecture of the Symbol Grounding Network (SGN), showing how object and predicate modules combine to produce probabilistic symbolic states.

Continuous Planner (CP)

The core innovation is the Continuous Planner, which redefines discrete symbolic planning steps—precondition checking, action selection, and state updates—as operations on probabilistic distributions. By treating the state as a distribution rather than a set of binary facts, the CP can "smooth over" perceptual errors. If the SGN predicts a state with 70% confidence, the planner can still compute a path toward the most likely goal configuration, rather than crashing due to an "invalid" discrete state.

Experimental Evaluation

We evaluated our framework in the BulletPhysics simulator across two domains:

  1. Block Stacking: Stacking 8 blocks into a target configuration.
  2. Object Sorting: Categorizing and moving objects based on a single demonstration.

We compared our approach against Neural Task Graphs (NTG), a prominent deep-learning baseline.

Key Results

  • Data Efficiency: In object sorting, our Continuous Planner achieved a 100% success rate with only 8 training tasks, whereas the NTG baseline failed to reach this benchmark even with double the training data.
  • Robustness: The CP significantly outperformed standard symbolic planners in block stacking. While traditional planners failed when SGN outputs were inconsistent, the CP successfully navigated these "noisy" states to find valid action sequences.
  • Zero-Shot Generalization: The use of a planner allows the system to handle environmental changes not seen in the demonstration. If an object is moved, the planner simply re-calculates the optimal path to the symbolic goal.

Discussion and Conclusion

The primary insight of this work is that disentangling perception from planning through modularity and continuous relaxation leads to massive gains in generalization and efficiency. By reusing "Predicate Modules," the system learns fundamental concepts (like "clearness") that apply across all tasks.

However, limitations remain. The current planner assumes deterministic actions and requires a predefined PDDL-style domain. Future work will focus on extending this framework to stochastic environments and raw pixel inputs, further reducing the need for prior domain knowledge while maintaining the benefits of structured planning.

Summary updated

Abstract

We address one-shot imitation learning, where the goal is to execute a previously unseen task based on a single demonstration. While there has been exciting progress in this direction, most of the approaches still require a few hundred tasks for meta-training, which limits the scalability of the approaches. Our main contribution is to formulate one-shot imitation learning as a symbolic planning problem along with the symbol grounding problem. This formulation disentangles the policy execution from the inter-task generalization and leads to better data efficiency. The key technical challenge is that the symbol grounding is prone to error with limited training data and leads to subsequent symbolic planning failures. We address this challenge by proposing a continuous relaxation of the discrete symbolic planner that directly plans on the probabilistic outputs of the symbol grounding model. Our continuous relaxation of the planner can still leverage the information contained in the probabilistic symbol grounding and significantly improve over the baseline planner for the one-shot imitation learning tasks without using large training data.
BibTeX
@inproceedings{huang2019continuous,
  author = {De-An Huang and Danfei Xu and Yuke Zhu and Animesh Garg and Silvio Savarese and Li Fei-Fei and Juan Carlos Niebles},
  title = {{Continuous Relaxation of Symbolic Planner for One-Shot Imitation Learning}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=PdH9c2QixEg},
}
Figure