In classical control theory and machine learning, we often assume the world is smooth. We rely on gradients, differential equations, and continuous kernels because they allow us to predict the future by looking at the immediate past. However, for a robot attempting to twirl a pen or walk across a room, the world is anything but smooth.
When a robotic finger strikes an object or a foot hits the pavement, the physics of the system changes instantaneously. These moments, known as hybrid dynamics, represent a fundamental challenge: how do you model and control a system where the rules of motion switch abruptly?
The Failure of Smoothness Priors
Most robotic models are built on "smoothness priors." If a robot moves its arm slightly, we expect the resulting state to change slightly. But as Marc Toussaint and Sethu Vijayakumar point out in Learning discontinuities with products-of-sigmoids for switching between local models, physical interactions like hitting a joint limit or touching a surface create sharp discontinuities.
Standard Gaussian kernels—the bread and butter of many learning algorithms—struggle here. They try to "smooth over" the impact, leading to models that are inaccurate exactly where they matter most: at the point of contact. To solve this, researchers are moving toward models that can represent "sharp" switches. Toussaint and Vijayakumar proposed using a "Product-of-Sigmoids" (PoS) architecture, which allows local models to switch abruptly, capturing the instantaneous change in dynamics without the interference issues found in traditional mixture models.
Why Contact-Rich Tasks are Hybrid
The difficulty of tasks like in-hand manipulation—reorienting an object within a grasp—stems from the fact that they are composed entirely of these discrete transitions. As Krishnan Srinivasan and Jeannette Bohg describe in Learning Hierarchical Control for Robust In-Hand Manipulation, every time a finger makes, breaks, or slides along a contact point, the underlying physics "jumps."
This creates a "non-smooth" landscape that breaks traditional optimization. If you are trying to plan a path using gradients, a sudden jump in force from zero to a high value (as occurs during contact) acts like a wall, making it impossible for standard solvers to "see" a path through the interaction. Yongpeng Jiang and Masayoshi Tomizuka argue in Robust Model-Based In-Hand Manipulation with Integrated Real-Time Motion-Contact Planning and Tracking that this is why model-based methods often stumble; they cannot accurately predict object behavior across these contact transitions in real-time.
Contact Implicit Control
One way to handle these jumps is to stop treating contact as a special "event" that needs a manual logic switch and instead bake it into the physics engine itself. This is the core intuition behind Contact-Implicit Model Predictive Control (CI-MPC).
In Linear Contact-Implicit Model-Predictive Control, Simon Le Cleac’h and Taylor A. Howell demonstrate a framework that allows robots to "discover" contact sequences. Instead of a programmer telling the robot "if you hit a wall, push against it," the controller reasons about the non-smooth forces directly. By using differentiable solvers that can handle the sudden transition of contact forces, the robot can autonomously decide to brace itself against a surface to maintain balance.
Bridging the Gap Between States
The challenge isn't just modeling the impact, but also the transition between different modes of existence—like a robot moving from free-space flight to ground contact. Xiaobin Xiong explores this in the context of "flying-hopping" locomotion in Reference-Steering via Data-Driven Predictive Control for Hyper-Accurate Robotic Flying-Hopping Locomotion. Here, the robot must maintain "hyper-accuracy" while switching between flight and ground phases.
Similarly, Mohan Sridharan and Saif Sidhik highlight the "changing-contact" problem in Towards a Framework for Changing-Contact Robot Manipulation. They emphasize that maintaining stability during the transition from free-space motion to surface interaction requires a framework that can handle the abrupt shift in physical properties.
The Limits of Stability Proofs
The non-smooth nature of contact doesn't just break our models; it breaks our math. Many theoretical guarantees for robot stability rely on the ability to calculate stabilizing gains using smooth dynamics. As Adam Block and Russ Tedrake note in Imitating Complex Trajectories: Bridging Low-Level Stability and High-Level Behavior, these proofs often fail in contact-rich scenarios. When objects collide, the standard control equations no longer apply, making it difficult to prove that a robot won't spin out of control during a complex assembly task.
Even advanced reinforcement learning struggles with this. Vittorio Giammarino and Ahmed H Qureshi point out in Goal Reaching with Eikonal-Constrained Hierarchical Quasimetric Reinforcement Learning that while physics-informed constraints (like the Eikonal equation) can help "stitch" together data for navigation, adapting these smooth partial differential equations to handle the sudden state transitions of contact remains a major open frontier.
Ultimately, the goal of this research direction is to move beyond the "smooth world" assumption. By developing controllers and models that embrace the discontinuous, hybrid nature of physical interaction, we can enable robots to perform the kind of dexterous, high-impact tasks that humans do every day.