Abstract
Deep Neural Networks (NNs) have been widely utilized in contact-rich manipulation tasks to model the complicated contact dynamics. However, NN-based models are often difficult to decipher which can lead to seemingly inexplicable behaviors and unidentifiable failure cases. In this work, we address the interpretability of NN-based models by introducing the kinodynamic images. We propose a methodology that creates images from the kinematic and dynamic data of a contact-rich manipulation task. Our formulation visually reflects the task's state by encoding its kinodynamic variations and temporal evolution. By using images as the state representation, we enable the application of interpretability modules that were previously limited to vision-based tasks. We use this representation to train Convolution-based Networks and we extract interpretations of the model's decisions with Grad-CAM, a technique that produces visual explanations. Our method is versatile and can be applied to any classification problem using synchronous features in manipulation to visually interpret which parts of the input drive the model's decisions and distinguish its failure modes. We evaluate this approach on two examples of real-world contact-rich manipulation: pushing and cutting, with known and unknown objects. Finally, we demonstrate that our method enables both detailed visual inspections of sequences in a task, as well as high-level evaluations of a model's behavior and tendencies.
Introduction
Deep Neural Networks (NNs) are excellent at modeling the messy, non-linear physics of "contact-rich" manipulation—tasks where a robot constantly interacts with objects, such as the friction of pushing a box or the varying resistance of slicing a carrot. However, these models are notoriously opaque. If a robot suddenly stops cutting because it predicts a failure, a human operator often has no idea why. This lack of interpretability is a major hurdle for deploying robots in safety-critical or collaborative environments. This paper introduces kinodynamic images, a novel way to represent robot sensor data as 2D images, allowing researchers to "peek inside" the black box of neural networks.
Methods: The Kinodynamic Image Pipeline
The core innovation is the transformation of abstract time-series vectors into a visual grid. Instead of feeding raw data into a network, the researchers arrange the data into a grid where the horizontal axis represents time and the vertical axis represents different physical features (e.g., change in position $\Delta p$, sensed forces $F$, and control inputs $u$).
By treating robot state data as an image, the authors can use standard Convolutional Neural Networks (CNNs) and apply Grad-CAM (Gradient-weighted Class Activation Mapping). Grad-CAM produces a "saliency map"—a heatmap overlaid on the input image—that highlights which specific features and which moments in time were most influential for a specific prediction, such as predicting when a knife is about to get stuck.
Experimental Evaluation
The researchers evaluated their approach on two real-world robotic tasks using an ABB YuMi arm:
- Pushing Task: 48 datasets involving pushing a box of varying weights (190g to 645g) across five different surfaces (cork, sandpaper, felt, etc.). The goal was to predict if the object would continue moving.
- Cutting Task: 43 trials of cutting vegetables (zucchinis, potatoes, and carrots) with varying controller stiffness. The goal was to predict "motion inhibition" (when the knife gets trapped).
Results and Interpretability
The models achieved high predictive accuracy, with F1-scores reaching up to 0.97 for pushing and roughly 0.82–0.85 for cutting tasks. However, the primary value of the kinodynamic representation is the ability to perform visual debugging. Unlike methods that compress data into an abstract latent space, kinodynamic images keep every sensor channel distinct. This ensures that when the model points to a "hot" pixel, the researcher knows exactly which physical property it corresponds to.
As shown in the saliency maps above, researchers can observe "label transitions." For example, as a robot transitions from successfully cutting to being trapped, the Grad-CAM heatmaps show the network's attention shifting from displacement sensors to force sensors. This analysis revealed that models with similar accuracy can rely on completely different physical cues; some models correctly prioritize downward progress, while others may be distracted by irrelevant lateral force noise.
Conclusion
Kinodynamic images provide a versatile framework for interpreting the decisions of neural networks in robotic manipulation. While currently limited to classification problems, the method enables detailed visual inspections of task sequences and high-level evaluations of model tendencies. Future work aims to extend this to regression tasks and real-time integration, allowing robots to explain their "intent" or potential failures to human collaborators during a task.
Summary updated