Standard neural networks are built on the assumption of smoothness. They excel at approximating continuous functions where a small change in input leads to a small change in output. However, physical reality is often "jumpy." When a billiard ball strikes a rail or a robotic gripper makes contact with a table, the system’s velocity changes near-instantaneously.
These events, known as discontinuous contact dynamics, represent a fundamental challenge for machine learning. Because standard regressors have an inherent "smoothness bias," they struggle to represent the sharp "jolts" of a collision, often resulting in physically impossible predictions like objects "ghosting" through one another or losing energy for no reason.
The Failure of Smoothness
In a typical neural world model, the network tries to map the current state to the next state. When this model encounters a collision, it sees a massive, sudden change in velocity. As Ruifeng Chen and colleagues explain in Offline Transition Modeling via Contrastive Energy Learning, attempting to fit these sharp transitions with standard regression creates high gradients that degrade the model’s ability to generalize. The network essentially tries to "round off" the corner of the collision, leading to a smoothed-over trajectory that violates the laws of physics.
This problem isn't limited to simple robotics. Even in computer vision, Qi Ma and Luc Van Gool have noted in Continuous Pose for Monocular Cameras in Neural Implicit Representation that neural networks used as smoothness priors can fail to capture high-frequency mechanical shocks or abrupt motion changes.
Embedding Physical Structure
To solve this, researchers have moved away from "black-box" models toward "gray-box" architectures that build physical constraints directly into the neural network.
One prominent approach is the use of Lagrangian mechanics. Instead of learning to predict velocity directly, models like Deep Lagrangian Networks (DeLaN), developed by Michael Lutter and Jan Peters, learn the underlying energy of the system. By forcing the network to obey the Euler-Lagrange equations, the model naturally conserves energy.
However, standard Lagrangian models still assume smoothness. To handle impacts, Andreas Hochlehnert and Marc Peter Deisenroth introduced CD-Lagrange networks in Learning Contact Dynamics using Physically Structured Neural Networks. This architecture incorporates physics-based integration schemes that specifically account for momentum transfer during impacts, allowing the network to represent sharp velocity changes without "smearing" them.
Implicit Representations and Energy Landscapes
Another way to bypass the smoothness bias is to stop asking the network to predict the exact next state. Instead, researchers use implicit modeling.
In ContactNets, Samuel Pfrommer and Michael Posa suggest that the neural network should only learn smooth geometric representations, such as Signed Distance Functions (SDFs). By delegating the "sharp" part of the motion—the actual impact—to established physical laws and solvers, the model avoids the numerical headaches of trying to learn a discontinuous function.
Similarly, the Energy-based Transition Model (ETM) proposed by Ruifeng Chen treats dynamics as an energy landscape. Instead of a regressor that maps $X \to Y$, the model scores candidate transitions. This allows the model to represent abrupt changes in dynamics because the underlying neural network is scoring the validity of a state rather than trying to approximate a discontinuous jump in motion.
Hybrid Symbolic and Iterative Approaches
A third direction involves combining neural learning with discrete logic or traditional solvers.
- Symbolic Hybrids: Joseph Z. Xu and John E. Laird demonstrated in Learning Integrated Symbolic and Continuous Action Models for Continuous Domains that symbolic logic can be used to identify different "modes" of behavior (e.g., "in flight" vs. "colliding"). By switching between different continuous models based on these symbolic states, the system can handle discontinuities that would baffle a single regressor.
- Solver Hybrids: In Accurately Solving Rod Dynamics with Graph Learning, Han Shao and colleagues used Graph Networks to provide a high-quality "initial guess" for a traditional iterative physics solver. This COPINGNet approach gets the speed of deep learning while relying on the solver to handle the hard physical constraints of contact and friction.
By moving away from pure end-to-end regression and toward these structured, implicit, or hybrid frameworks, researchers are creating world models that can finally "feel" the hard edges of the physical world.