Bridging the Kinodynamic Feasibility Gap in Robot Navigation

30 concepts31 papers2 proposals

The classical robotic navigation pipeline splits autonomy into two distinct problems: a high-level geometric planner that finds a collision-free path through an environment, and a low-level controller that calculates the motor torques, steering angles, or thrust commands needed to track that path.

While this hierarchical separation makes the math tractable, it introduces a fundamental flaw: high-level planners are typically dynamics-agnostic. They treat robots as simplified geometric shapes—like omnidirectional points or disks—and assume the system can instantaneously change direction or accelerate. When a real physical system attempts to execute these idealized paths, it encounters a kinodynamic feasibility gap (or kinematic-dynamic mismatch), where the planned path violates the true physical capabilities of the hardware.

The Cost of Decoupling Planning from Control

When a high-level planner operates without an understanding of low-level physics, the resulting trajectories often demand unachievable maneuvers. This mismatch manifests across diverse robotic domains:

  • Inertia and Underactuation: An unmanned underwater vehicle might drift far from its intended path due to hydrodynamic forces and momentum (Lee et al.), while a multi-rotor drone transporting a slung payload may experience motor saturation or severe tracking errors when trying to execute sudden accelerations (De Croon et al.; Wahba et al.).
  • Nonholonomic Constraints: For car-like robots, standard grid-based paths ignore Ackermann steering geometry. When a controller attempts to smooth these discrete paths, the actual trajectory deviates significantly, risking unexpected collisions (Wen et al.).
  • Whole-Body Stability: In legged quadrupeds and humanoids, a geometric planner might place a target waypoint directly over a deep gap or an unstable surface. This forces the low-level controller to attempt maneuvers that compromise balance or cause mechanical failure (Wang et al.; Xiao et al.; Truong et al.).

As noted by Zheng et al., this two-stage approach inevitably struggles in fast-changing environments. Because the high-level planner is too slow to recompute paths and the low-level controller is blind to global safety constraints, the robot is forced to track an unexecutable trajectory, leading to unstable behavior or collisions (Xiong et al.; Tajbakhsh et al.).

To close this gap, modern research focuses on tightly coupling planning and control. Rather than treating them as isolated steps, contemporary frameworks infuse physical dynamics directly into the planning phase using three core strategies: physics-aware search, policy-aware filtering, and differentiable co-design.

Infusing Physics into Search and Optimization

The most direct way to eliminate the feasibility gap is to replace simple geometric distance metrics in the planner with cost functions derived from the robot's actual differential equations of motion.

Instead of planning a geometric path and smoothing it later, Atanasov et al. solve a Linear Quadratic Minimum Time (LQMT) control problem to generate a highly accurate search heuristic for quadrotors. This ensures that the search tree only expands along trajectories that are natively executable within the vehicle's thrust and time limits. For multi-robot systems, Tajbakhsh et al. combine Conflict-Based Search with Model Predictive Control (MPC), introducing collision constraints dynamically so that the optimization natively respects actuation limits without sacrificing scalability.

[Traditional Pipeline]
Geometric Planner  ──(Infeasible Waypoints)──>  Low-Level Controller  ──> Tracking Failure

[Physics-Aware Search]
LQMT / MPC Heuristics + Kinodynamic Constraints  ──>  Directly Executable Trajectory

This integration is equally vital in specialized domains. In spacecraft navigation, the authors of Simultaneous Control and Trajectory Estimation for Collision Avoidance of Autonomous Robotic Spacecraft Systems unify control, planning, and estimation into a single probabilistic factor graph, ensuring thruster constraints are never violated during collision avoidance. Even in coverage tasks like ergodic search, ignoring physical obstacles during global optimization forces a decoupled local planner to make inefficient corrections; integrating these constraints directly preserves global optimality (Ren et al.).

Policy-Aware Planning and Feasibility Filters

When exact mathematical models of a robot's physics are too complex, non-linear, or noisy to plan with directly, learning-based abstractions can bridge the gap.

One approach is to make global planners policy-aware. In the PRM-RL framework developed by Faust et al. and expanded by Francis et al., a global Probabilistic Roadmap does not assume straight-line connectivity between nodes. Instead, a local reinforcement learning (RL) agent attempts to navigate between sampled nodes in a physics simulator. If the local policy fails due to the robot's physical constraints or sensor noise, that edge is rejected from the roadmap. This ensures that the final graph contains only paths that the low-level controller is proven capable of executing.

Similarly, Juvvi et al. use RL to master basic motion primitives, which a sampling-based planner then stitches together to satisfy complex temporal logic rules without needing an explicit analytical model of the robot's physics.

[Policy-Aware Planning (PRM-RL)]
Sample Nodes  ──>  Local RL Controller Tests Edge  ──>  If Successful  ──> Add to Roadmap
                                                   ──>  If Fails       ──> Reject Edge

Alternatively, learning can act as a vehicle-specific filter. Liu et al. repurpose RL value functions as a feasibility filter. When a generic, vehicle-agnostic behavior model proposes a sequence of waypoints, the value function evaluates whether a specific chassis (e.g., a heavy truck versus a nimble sedan) can physically execute the maneuver, refining the path to match the hardware's limits. For multi-agent forecasting, Ivanovic represents predicted trajectories as linear-affine dynamical systems, allowing planners to directly optimize control inputs rather than relying on discrete waypoint sampling.

Co-Design and Differentiable Pipelines

The ultimate realization of this research direction is the complete unification of planning and control into a single, differentiable optimization loop. By making the low-level controller differentiable, the high-level planner can receive gradients that describe exactly why a proposed path was difficult to track.

The iWalker system by Xiao et al. demonstrates this for humanoid robots. By using Imperative Learning, they embed a physics-based MPC directly into the training loop of a visual planning network. Because the entire pipeline is differentiable, the visual planner is forced to learn a representation that only proposes paths the low-level MPC can physically stabilize.

Similarly, Xiong et al. present a co-design framework where high-level path planning and low-level control policies are optimized jointly under formal logic specifications, drastically reducing tracking errors.

To make these joint optimization frameworks computationally viable for real-time control, researchers often leverage historical data. As demonstrated by Lembono, a "Memory of Motion" regression model can warm-start optimal control solvers, reducing the computational burden of finding dynamically optimal trajectories by up to 70%. By combining the predictive power of optimization with the speed of learned representations, robots can finally navigate complex environments without outrunning their own physics.

Key Foundational Work

Analogical Path Planning (2004) — contributed 1 concept elements to this direction.
Search-based Motion Planning for Quadrotors using Linear Quadratic Minimum Time Control (2017) — contributed 1 concept elements to this direction.
A Algorithm-Based Practical Path Planning Considering the Actual Dynamic Behavioural Constraint in Unmanned Underwater Vehicles (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_kinematic_dynamic_mismatch_81fda144,
  title = {kinematic-dynamic mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/kinematic-dynamic-mismatch/rdr_81fda14408ac492185b19d44f1f7bb3e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_kinematic_dynamic_mismatch_81fda144_v1,
  title = {kinematic-dynamic mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/kinematic-dynamic-mismatch/rdr_81fda14408ac492185b19d44f1f7bb3e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake