[Paper Review] ContactNets: Learning Discontinuous Contact Dynamics with Smooth, Implicit Representations

Samuel Pfrommer, Mathew Halm+1 more|arXiv (Cornell University)|Sep 23, 2020

Robot Manipulation and Learning18 citations

TL;DR

ContactNets introduces a differentiable, implicit neural representation for learning discontinuous contact dynamics in robotics by parameterizing inter-body signed distance and contact-frame Jacobians. It achieves realistic impact, non-penetration, and stiction behavior with only 60 seconds of real-world data, outperforming end-to-end models in physical plausibility and generalization by enforcing complementarity and maximum dissipation via a novel loss function.

ABSTRACT

Common methods for learning robot dynamics assume motion is continuous, causing unrealistic model predictions for systems undergoing discontinuous impact and stiction behavior. In this work, we resolve this conflict with a smooth, implicit encoding of the structure inherent to contact-induced discontinuities. Our method, ContactNets, learns parameterizations of inter-body signed distance and contact-frame Jacobians, a representation that is compatible with many simulation, control, and planning environments for robotics. We furthermore circumvent the need to differentiate through stiff or non-smooth dynamics with a novel loss function inspired by the principles of complementarity and maximum dissipation. Our method can predict realistic impact, non-penetration, and stiction when trained on 60 seconds of real-world data.

Motivation & Objective

• To address the fundamental mismatch between smooth neural network inductive biases and the discontinuous, non-smooth nature of frictional contact in robotics dynamics.
• To enable data-efficient learning of realistic contact behaviors—such as impact, stiction, and non-penetration—without requiring force or contact sensors.
• To reformulate contact dynamics learning as a smooth, implicit optimization problem compatible with existing simulation, control, and planning pipelines.
• To overcome the numerical instability of backpropagating through discontinuous simulators by introducing a novel loss grounded in contact mechanics principles.
• To demonstrate that implicit, physics-aware neural parameterizations can outperform standard end-to-end models in both accuracy and physical consistency.

Proposed method

• Represents contact dynamics implicitly using signed distance functions between rigid bodies and contact-frame Jacobians, enabling smooth optimization despite discontinuous physics.
• Parameterizes the system using a deep neural network to predict the height of the contact surface and contact Jacobians, avoiding direct regression of discontinuous forces.
• Employs a novel differentiable loss function inspired by the principles of complementarity and maximum dissipation, which stabilizes training and avoids ill-posed gradients.
• Uses gravitational acceleration as a prior to guide the network toward physically plausible contact states, improving generalization from limited data.
• Enables compatibility with standard robotics simulation and control stacks by using a representation equivalent to that used in physics engines (e.g., [24, 11, 21]).
• Trains end-to-end on real-world 3D manipulation data (60 seconds of tosses) without requiring explicit contact or force sensing.

Experimental results

Research questions

• RQ1 Can a smooth, implicit neural representation effectively model discontinuous contact dynamics in robotics without explicit contact sensing?
• RQ2 How can a differentiable loss function be designed to stabilize training when backpropagating through non-smooth contact events?
• RQ3 Can implicit parameterization of signed distance and contact Jacobians lead to better generalization and physical plausibility than end-to-end models?
• RQ4 To what extent does incorporating contact mechanics principles (complementarity, maximum dissipation) improve model performance on real-world data?
• RQ5 How does the model perform in capturing non-penetration, impact, and stiction behaviors under limited training data?

Key findings

• ContactNets achieved at least comparable positional accuracy and superior rotational accuracy to a 256-toss end-to-end model after only 32 training tosses.
• ContactNets Polytope averaged only 1.9% block width penetration, significantly lower than the end-to-end model’s over 20% penetration.
• ContactNets rollouts exhibited minimal penetration and realistic discontinuous impulses, while end-to-end models frequently drifted and failed to stop after contact.
• The model successfully captured hard face-to-ground contacts and rotational stability, even without explicit knowledge of ground height or orientation.
• The method demonstrated strong data efficiency, producing physically plausible dynamics from just 60 seconds of real-world data.
• ContactNets outperformed the end-to-end baseline in both positional and rotational error, especially with ample training data, due to its implicit enforcement of physical constraints.