In traditional robotics, movement is often treated as a choreography of pre-planned events. To make a robot walk, an engineer typically writes a "contact schedule"—a script that tells the robot exactly when each foot should strike the ground and for how long. While this works for steady walking on flat floors, it fails in the messy, unpredictable real world. If a robot is pushed or encounters a gap, a fixed schedule becomes a liability.
The research direction emerging from these papers moves away from scripted schedules toward Contact-Implicit Trajectory Optimization (CIO). The goal is to allow robots to autonomously "discover" when and where to make contact with their environment to achieve a goal, treating the contact sequence itself as a variable to be optimized rather than a constraint to be followed.
The Contact Schedule Bottleneck
Most motion planning systems, as noted by Diego Pardo and Jonas Buchli in Projection based whole body motion planning for legged robots, depend on these manually predefined sequences. This reliance prevents "reactive stepping"—the ability of a robot to instinctively move its foot to catch a fall.
The mathematical reason for this bottleneck is that contact is "non-smooth." In physics, the moment a foot touches the ground, the forces jump from zero to hundreds of Newtons instantaneously. Standard optimization algorithms, which rely on smooth gradients to find the "downhill" path to a solution, get stuck on these sharp discontinuities. To avoid this, researchers traditionally split the problem: humans decide the discrete contact sequence, and the computer optimizes the smooth motion within those contacts.
Smoothing the Physics Landscape
To break this bottleneck, researchers like Quentin Le Lidec and Justin Carpentier have introduced Randomized Differential Dynamic Programming (RDDP). In Leveraging randomized smoothing for optimal control of nonsmooth dynamical systems, they apply "randomized smoothing" to the physics of friction and impact.
By adding a small amount of controlled noise to the dynamics during the planning phase, the "jagged" optimization landscape of impacts becomes a smooth hill. This allows gradient-based solvers to "see" through the impact and understand how changing a movement now will affect a contact that hasn't happened yet. This enables robots to solve tasks like jumping or lifting objects without being told the timing in advance.
Contact as a Complementarity Constraint
Another powerful approach treats contact as a mathematical "complementarity constraint." This is the logic used by Arvind Raghunathan in the PYROBOCOP framework and Haokun Wang in Contact-Aware Motion Planning Among Movable Objects.
The intuition is simple: either the distance between a foot and the ground is greater than zero (no contact), or the force between them is greater than zero (contact). They cannot both be non-zero at the same time. By formulating contact this way, the optimizer can decide to "turn on" a force whenever it helps reach the goal. This allows a robot to dynamically decide to push a box out of its way or brace itself against a wall to maintain balance, as demonstrated in the CI-MPC framework by Simon Le Cleac’h and Mac Schwager in Linear Contact-Implicit Model-Predictive Control.
Real-Time Discovery of Agile Maneuvers
For these ideas to work on real hardware, they must be fast. Carlos Mastalli and colleagues developed Crocoddyl, which uses Feasibility-driven Differential Dynamic Programming (FDDP). This algorithm allows robots to plan highly dynamic maneuvers, like front-flips, in milliseconds.
When the optimization is fast enough, the robot no longer needs a "gait." Instead, the movement emerges from the task. In Contact-Implicit Trajectory Optimization for Dynamic Object Manipulation, Jan Carius and Marco Hutter show that a robot can discover whether it should use a quick strike or a sustained push to move an object, simply by optimizing for the end goal.
Similarly, in the realm of dexterous manipulation, Yongpeng Jiang and Mingrui Yu demonstrated in Robust Model-Based In-Hand Manipulation that jointly optimizing finger trajectories and contact forces allows complex behaviors like "finger gaiting" to emerge naturally. The robot "learns" to walk its fingers around an object because that is the most efficient way to rotate it, not because it was programmed with a finger-stepping sequence.
From Periodic Gaits to Autonomous Exploration
The shift toward contact-implicit methods represents a move from periodic locomotion (repeating a cycle) to non-periodic locomotion. As Ilyass Taouil and Lorenzo Amatucci argue in Non-Gaited Legged Locomotion with Monte-Carlo Tree Search, treating gait selection as a search problem allows robots to react to heavy hits and uneven terrain by placing feet exactly where they are needed, regardless of the "rhythm."
The ultimate goal, as framed by Hae-Won Park 0002 in Dynamic Policy Learning for Legged Robot, is to enable robots to discover optimal contact points and timings entirely through exploration. By removing the "crutch" of human-engineered contact schedules, we allow robots to exploit the full range of their physical capabilities, leading to the kind of agility seen in parkour or complex multi-fingered manipulation.