Combining learned and analytical models for predicting action effects from sensory data

Alina Kloss, Stefan Schaal, Jeannette BohgInt. J. Robotics Res 2022 · OpenReview

TL;DR

This paper introduces a hybrid approach that combines physics-based analytical models with neural networks for predicting the effects of physical interactions (like pushing) from raw sensory data. This hybrid method significantly reduces the amount of training data needed and improves generalization to new scenarios compared to purely learned models, while still leveraging the strengths of deep learning for perception.

What Problem Does This Paper Address?

Robots need to predict how their actions will affect the physical world to make intelligent decisions. Traditionally, this has been done using analytical physics models. However, these models often require precise, hard-to-obtain state information (like object shape and precise contact points) from raw sensor data. More recent approaches use neural networks to learn these dynamics directly from sensory input, but they often struggle to generalize beyond their training data and lack interpretability. This paper addresses the challenge of bridging this gap: how to leverage the interpretability and generalization capabilities of analytical models with the perceptual power of neural networks, especially when dealing with complex, real-world interactions like planar pushing.

Methods

The core contribution is a hybrid architecture that uses a neural network for perception and an analytical model for prediction.

  • Perception Network: A convolutional neural network (CNN) takes raw depth images (or point clouds) as input. It extracts a "glimpse" (a small image patch around the robot's end-effector) to focus on relevant interaction details and also estimates the object's overall position. This processed information is then fed into the prediction module.
  • Analytical Prediction Model: The paper uses a well-established analytical model for quasi-static planar pushing. This model predicts the object's linear and angular velocity based on inputs like the effective push velocity, contact point, and surface normal.
  • Hybrid Integration: The neural network's perception output is used to provide the necessary inputs (like contact point and surface normal) for the analytical model. This allows the system to learn from raw sensory data while benefiting from the physics-based reasoning.
  • Variants Explored:
    • Neural Network Only (neural): A purely learned approach where a neural network handles both perception and dynamics prediction.
  • Full Analytical Model (hybrid): The proposed approach using a CNN for perception and the full analytical model for prediction.
  • Simplified Analytical Model (simple): A hybrid approach where the neural network predicts the "effective push velocity" (an intermediate step in the analytical model), and the rest of the analytical model is used.
  • Full Analytical Model + Error Term (error): A hybrid model where an additional learned "error-correction" term is added to the output of the analytical model to compensate for its potential inaccuracies.

Datasets

The primary dataset used is the MIT Push Dataset. This dataset contains recordings of planar pushing experiments with various objects and surfaces, including object pose and force data. For this work, the authors rendered depth images from this dataset to create a large-scale benchmark. They also augmented the dataset with simulated images and contact information.

The experiments involved creating specific training and testing splits, including:

  • A large dataset for evaluating data efficiency (190k training, 38k testing).
  • Subsets of the training data to test data efficiency with varying amounts of training examples (from 2500 to 190k).
  • Held-out push configurations (angles, contact points) for generalization tests.
  • Datasets with different push velocities for extrapolation tests.
  • Datasets with unseen object shapes for generalization tests.

Key Insights and Ideas

  • Hybrid is Better for Data Efficiency and Generalization: The core finding is that combining learned perception with analytical dynamics significantly outperforms purely learned models when training data is scarce or when generalizing to unseen scenarios (new push angles, velocities, or objects).
  • Analytical Models Provide Inductive Bias: The analytical model acts as a strong inductive bias, guiding the learning process and preventing the model from overfitting to the training data. This is crucial for generalization.
  • Interpretability Advantage: The hybrid approach offers better interpretability because the intermediate representations (like contact point and normal) have physical meaning, unlike the abstract representations learned by purely neural models.
  • Error Correction for Analytical Model Limitations: The "error" variant, which adds a learned correction term to the analytical model's output, shows promise in compensating for inaccuracies in the analytical model itself, especially when the model's assumptions are violated.
  • Pure Neural Networks Can Excel with Enough Data: When trained on a very large dataset and evaluated within its training domain, a purely neural network can achieve the highest accuracy, suggesting that it can learn complex dynamics if given sufficient examples. However, this advantage disappears when generalization is tested.

Key Results

  • Data Efficiency: The hybrid approach (hybrid and error) achieved comparable performance to the physics baseline with only 2500 training examples, whereas purely learned models required over 20,000 examples.
  • Generalization: In tests with unseen push angles and contact points, the hybrid and error models performed best, outperforming purely learned models.
  • Extrapolation to Velocities: When extrapolating to higher push velocities not seen during training, the hybrid and error models showed significantly better performance than purely learned models, staying much closer to the physics baseline.
  • Generalization to Objects: The hybrid approach also facilitated generalization to novel object shapes, which was more challenging for purely learned models.
  • Pure Neural Network Performance: With a full dataset (190k examples), the purely neural network performed best within its training domain, even outperforming the physics baseline. However, this advantage vanished when generalization was tested.
  • Error Correction Impact: The error-correction term improved the hybrid model's ability to compensate for errors in the analytical model, especially when the friction parameter was manipulated.

Limitations and Future Work

  • Analytical Model Dependency: The hybrid approach relies on the availability of a suitable analytical model. Finding differentiable analytical models for complex physical processes can be challenging.
  • Switching Dynamics: The paper notes that handling switching dynamics (e.g., transitions between sticking and sliding contact) remains a challenge.
  • Non-Trivial Viewpoints: While the hybrid approach was extended to work with non-top-down camera views, it requires known camera parameters and transforms, and training becomes more challenging.
  • Future Work: The authors suggest extending the work to more complex perception problems (RGB images, multiple objects), using point cloud networks, predicting multiple steps into the future, and incorporating learned models into model predictive control. They also highlight the potential of learning predictive models with neural networks when analytical models are unavailable.

Summary updated

Abstract

One of the most basic skills a robot should possess is predicting the effect of physical interactions with objects in the environment. This enables optimal action selection to reach a certain goal state. Traditionally, dynamics are approximated by physics-based analytical models. These models rely on specific state representations that may be hard to obtain from raw sensory data, especially if no knowledge of the object shape is assumed. More recently, we have seen learning approaches that can predict the effect of complex physical interactions directly from sensory input. It is, however, an open question how far these models generalize beyond their training data. In this work, we investigate the advantages and limitations of neural-network-based learning approaches for predicting the effects of actions based on sensory input and show how analytical and learned models can be combined to leverage the best of both worlds. As physical interaction task, we use planar pushing, for which there exists a well-known analytical model and a large real-world dataset. We propose the use of a convolutional neural network to convert raw depth images or organized point clouds into a suitable representation for the analytical model and compare this approach with using neural networks for both, perception and prediction. A systematic evaluation of the proposed approach on a very large real-world dataset shows two main advantages of the hybrid architecture. Compared with a pure neural network, it significantly (i) reduces required training data and (ii) improves generalization to novel physical interaction.
BibTeX
@article{kloss2022combining,
  author = {Alina Kloss and Stefan Schaal and Jeannette Bohg},
  title = {{Combining learned and analytical models for predicting action effects from sensory data}},
  journal = {Int. J. Robotics Res.},
  year = {2022},
  url = {https://openreview.net/forum?id=uXu6AdXAqd},
}
Figure