Planning for Learning Object Properties

Leonardo Lamanna, Luciano Serafini, Mohamadreza Faridghasemnia, Alessandro Saffiotti, Alessandro Saetti, Alfonso Gerevini, Paolo Traverso

TL;DR

This paper introduces a novel approach that uses symbolic planning to automate the process of training neural networks to recognize object properties. By framing the learning task as a planning problem, the agent can autonomously decide when and how to collect data and train its perception models, enabling robots to adapt to new environments without pre-existing labeled datasets.

What Problem Does This Paper Address?

Autonomous agents, like robots, need to understand their environment to achieve goals. A crucial part of this is recognizing objects and their properties (e.g., is a door open or closed?). Traditionally, this is done by training supervised machine learning models on massive, pre-labeled datasets. However, in real-world, open-ended scenarios, it's impossible to have a pre-trained model for every possible object and property. This paper tackles the challenge of enabling agents to autonomously learn these perceptual abilities online, as they interact with their environment, without human supervision. The core problem is bridging the gap between low-level sensory data (like images) and high-level symbolic representations used for planning, specifically for learning new object properties.

Methods

The paper's core innovation is framing the problem of learning object properties as a symbolic planning task. They extend a standard PDDL (Planning Domain Definition Language) planning domain with new predicates and actions specifically for learning. This "learning domain" allows the agent to:

  • Observe objects: Actions like Observe(o, t, p) are defined to collect sensory data (e.g., images) of an object o of type t related to property p.
  • Explore for new data: Actions like Explore_for(t, p) help the agent search for new objects or situations to gather more diverse training examples.
  • Train perception models: An action Train(t, p, q) triggers the training of a neural network ρt,p using collected data.

The agent plans using a combined domain (base + learning domains) to generate a sequence of actions. This plan guides the agent to interact with the environment, collect relevant data, and train its perception models. Crucially, the system accounts for the fact that real-world execution might not perfectly match the symbolic plan (e.g., an action to "go close" might not actually make the agent close enough). Therefore, after executing actions, the agent re-observes the environment and replans if necessary.

Datasets

The paper evaluates their approach using the ITHOR (Interactive 3D Environment for Visual AI) simulator. ITHOR provides a photo-realistic environment with various indoor scenes (kitchens, living rooms, etc.) and allows for object interaction.

  • Training/Validation: The authors generated training and validation sets for object detection using ITHOR's environments, resulting in a large dataset of object types and bounding boxes.
  • Testing: For evaluating property predictors, they generated ground truth test sets (Gt,p) by randomly navigating 20 testing environments within ITHOR. These sets contain object instances labeled with specific properties (e.g., Is_Open, Dirty, Toggled, Filled).
  • Real-world: For the real-world demonstration, they used a SoftBank Pepper robot. The object detector was a pre-trained YoloV5 model.

Key Insights and Ideas

  • Planning as a Learning Orchestrator: The most significant insight is using symbolic planning not just for task execution, but as a meta-level controller to orchestrate the entire learning process. This allows for a principled way to decide what to learn, when to learn it, and how to gather the necessary data.
  • Bridging Symbolic and Subsymbolic: The approach effectively bridges the gap between symbolic planning (abstract goals and actions) and subsymbolic perception (raw sensor data processed by neural networks). The planning domain provides the abstract goals (e.g., "learn property X"), and the learned neural networks provide the low-level perception capabilities needed to achieve those goals.
  • Autonomous Data Collection: The system autonomously decides which objects to observe and how to manipulate them (or ask a human to) to gather positive and negative examples for a property. This is a key step towards truly autonomous agents.
  • Robustness to Noisy Perception: The experiments comparing "Noisy Detections" (ND) with "Ground Truth Detections" (GTD) highlight that the approach can still learn effectively even with imperfect object detection, though performance is better with accurate detections.

Key Results

  • Simulated Environment: The approach demonstrated successful learning of object properties (Is_Open, Dirty, Toggled, Filled) across 32 object types.
  • Precision and recall metrics were evaluated. In general, the GTD version (using perfect object detection) achieved higher weighted average precision and recall than the ND version.
  • For the Dirty property, GTD precision was significantly higher than ND for most object types, though a notable exception was the bed object.
  • The recall was often lower than precision, attributed to imbalanced training datasets (more negative than positive examples).
  • Real World Demonstrator: Using a Pepper robot, the system learned properties like Is_Open and Filled with high precision and recall (e.g., weighted average precision of 0.88 and recall of 0.95 on tested pairs).
  • Comparison to Baseline: The learned predictors achieved precision comparable to an offline baseline trained on manually collected data, though recall was slightly worse.

Limitations and Future Work

  • Object Detector Dependency: The performance is significantly impacted by the quality of the pre-trained object detector. While the system can learn with noisy detections, perfect detections lead to better results.
  • Action Execution Uncertainty: The paper acknowledges that action execution might not perfectly achieve the intended symbolic state, requiring replanning. However, the focus is on learning properties, not necessarily on robust action execution itself.
  • Human Collaboration in Real World: In the real-world setup, the robot relies on a human to perform object manipulations.
  • Future Work: The authors suggest several promising directions:
  • Online Training of Object Detectors: Extending the system to learn or adapt the object detector itself.
  • Learning Relations: Learning relationships between different objects.
  • Improving Real-World Robustness: Adapting the approach for scenarios where the assumption of being able to change an object's property without fully recognizing it is more critical.
  • Learning from Scratch: Developing methods that don't rely on a pre-trained object detector at all.

Summary updated

Abstract

Autonomous agents embedded in a physical environment need the ability to recognize objects and their properties from sensory data. Such a perceptual ability is often implemented by supervised machine learning models, which are pre-trained using a set of labelled data. In real-world, open-ended deployments, however, it is unrealistic to assume to have a pre-trained model for all possible environments. Therefore, agents need to dynamically learn/adapt/extend their perceptual abilities online, in an autonomous way, by exploring and interacting with the environment where they operate. This paper describes a way to do so, by exploiting symbolic planning. Specifically, we formalize the problem of automatically training a neural network to recognize object properties as a symbolic planning problem (using PDDL). We use planning techniques to produce a strategy for automating the training dataset creation and the learning process. Finally, we provide an experimental evaluation in both a simulated and a real environment, which shows that the proposed approach is able to successfully learn how to recognize new object properties.
BibTeX
@inproceedings{lamanna2023planning,
  author = {Leonardo Lamanna and Luciano Serafini and Mohamadreza Faridghasemnia and Alessandro Saffiotti and Alessandro Saetti and Alfonso Gerevini and Paolo Traverso},
  title = {{Planning for Learning Object Properties}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i10.26416},
}
Figure