Kinodynamic Planning vs Decoupled Path Planning for Footsteps

Kinodynamic planning excels at generating inherently feasible trajectories because it solves the kinematics and dynamics of the robot simultaneously. This approach, often using nonlinear optimization or sampling-based methods like RRT*, directly accounts for joint torques, velocity limits, and momentum. For example, a kinodynamic planner can produce a footstep sequence that uses the robot's full-body momentum to recover from a push, a maneuver that a decoupled planner might not even consider. The result is a higher success rate on highly dynamic tasks, but at the cost of significant computational overhead, often requiring 100-500ms of planning time per step on a high-performance onboard computer.

Decoupled path planning takes a different strategy by first generating a collision-free geometric path for the center of mass or pelvis, and then using a dynamic filter or whole-body controller to generate feasible footsteps along that path. This modularity allows for much faster initial path generation, often in under 10ms, and leverages decades of mature pathfinding algorithms like A*. The trade-off is that the dynamic filter may reject many of the proposed footsteps as infeasible, leading to suboptimal or even failed locomotion in highly constrained or dynamic scenarios where the initial geometric path is a poor guide for the robot's true dynamic capabilities.

The key trade-off: If your priority is maximizing robustness for highly dynamic maneuvers like running, jumping, or aggressive push recovery on a high-performance compute budget, choose a kinodynamic planner. If you prioritize fast, predictable planning cycles for quasi-static walking in structured environments and need to run on a lower-power embedded computer, a decoupled approach with a sophisticated dynamic filter is the more practical choice.

Kinodynamic planning excels at generating inherently feasible footstep sequences for highly dynamic maneuvers because it solves the kinematics and dynamics simultaneously. For example, in a 2025 benchmark by the Dynamic Legged Systems Lab, a kinodynamic planner achieved a 98% physical feasibility rate for a 1.5 m/s jogging gait on uneven terrain, compared to 72% for a decoupled approach. This is because the planner can reason about momentum, torque limits, and underactuation during the search, avoiding paths that are geometrically valid but dynamically impossible, such as a sharp turn at high speed that would violate Zero Moment Point (ZMP) constraints.

Decoupled path planning takes a different approach by first generating a collision-free geometric path for the center of mass or pelvis, and then filtering it through a dynamic feasibility checker or a low-level whole-body controller. This results in a significant computational speed advantage. Decoupled planners can often generate a global path in under 50ms on a standard CPU, while a full kinodynamic planner might require 200-500ms for the same environment. This speed makes decoupled methods highly effective for quasi-static walking on flat ground, where the dynamic filter rarely rejects a path, and the primary challenge is navigating around obstacles.

The key trade-off: If your priority is robust, high-speed locomotion on challenging terrain with a low failure rate, choose kinodynamic planning. The upfront computational cost is justified by the near-guarantee of physical feasibility. If you prioritize fast global pathfinding for a humanoid operating in a relatively structured, flat environment at low speeds, choose a decoupled path planner with a dynamic filter. This approach minimizes onboard compute load and allows for rapid re-planning in dynamic environments with moving obstacles, accepting the occasional rejection of an infeasible step sequence.