Imagine trying to parallel park a car that can only move forward and backward, and whose steering wheel is locked. To get into the spot, you cannot simply slide sideways; you have to exploit the vehicle's momentum, gravity, or complex multi-step maneuvers. This scenario illustrates the core challenge of underactuated systems: mechanical systems that possess fewer independent control inputs (actuators) than degrees of freedom (DOFs).
In traditional robotics, controllers often assume full actuation—meaning every joint or axis of motion can be directly commanded by a motor. However, many real-world systems are inherently underactuated due to design constraints, weight reduction, or the laws of physics. Bipedal walking robots, quadrotors, cranes, and marine vessels all have fewer actuators than the total dimensions of their state space. Controlling them requires moving away from "forcing" a system along a path and toward intelligently steering its natural physical dynamics.
The Degree-of-Freedom Mismatch
To understand why underactuated control is difficult, consider a standard quadrotor. It operates in a six-dimensional space (three translational positions, three rotational orientations) but only has four independent rotor inputs. As a result, it cannot translate laterally without first tilting its body. This coupling between orientation and translation means that errors in one dimension cannot be corrected without temporarily disturbing another.
As noted by Del Prete et al. and Peters et al., standard robotic control frameworks optimized for fully actuated systems break down when applied to underactuated or floating-base platforms. In a fully actuated system, you can mathematically "invert" the dynamics: if you know your desired acceleration, you can directly calculate the required motor torques. In an underactuated system, this inversion is generally impossible because the dimensions of the control inputs and the output states do not match. As Kürkçü demonstrates in Algebraic Control: Complete Stable Inversion with Necessary and Sufficient Conditions, determining whether a specific output trajectory is even reachable becomes a highly complex algebraic problem when dealing with non-square control dimensions.
Energy Shaping and Geometric Control
If we cannot directly force an underactuated system into a desired configuration, the next best alternative is to manipulate its energy landscape so that it wants to go there naturally. This is the intuition behind energy shaping and Controlled Lagrangians.
In Controlled Lagrangians and Potential Shaping for Stabilization of Discrete Mechanical Systems, Bloch et al. show how control inputs can be used to mathematically alter the kinetic and potential energy properties of a system. Instead of fighting the unactuated dynamics, the controller reshapes the system's total energy function such that the desired target state becomes a natural minimum (a "well" that the system falls into).
Similarly, Fang et al. utilize geometric frameworks to design Proportional-Derivative (PID) controllers for cart-pendulum systems. By exploiting the inherent geometric coupling between the controlled cart and the uncontrolled pendulum, they achieve exponential stability around unstable equilibrium points without requiring full control authority over every component.
Exploiting Mathematical Structure: Flatness and Kinematics
For certain underactuated systems, clever mathematical transformations can reveal hidden pathways to simple control. One such property is differential flatness. A system is differentially flat if all of its states and control inputs can be expressed in terms of a carefully chosen set of "flat outputs" and their derivatives.
Lefebvre et al. leverage this property for slider-pusher systems (where a robot manipulates an object purely by pushing it). Because the system is underactuated, the robot cannot pull or directly steer the object in every direction. However, by treating the contact dynamics within a flat output space, the authors decouple complex path planning into a two-step process: first, planning a geometric route in the flat space, and second, optimizing the execution speed along that route.
When systems evolve on complex geometric spaces (like rotations in 3D space), closed-form kinematic planners can be derived. Cortes et al. developed a catalog of "switch-optimal" planners that calculate the exact sequences of maneuvers required to navigate underactuated systems on matrix groups, minimizing the number of times a system must stop or change directions to reach its goal.
Learning to Navigate the Mismatch
When a system's dynamics are too non-linear or poorly modeled to solve analytically—such as a bipedal robot walking on compliant, uneven terrain—modern data-driven methods offer an alternative approach.
Deep Reinforcement Learning (DRL) is highly effective at discovering coordinated policies that exploit underactuated coupling. For example, Xie et al. successfully trained a feedback controller for the Cassie bipedal robot. Because the robot's feet lack direct actuation against the ground, its balance depends entirely on unactuated physical interactions. The DRL framework bypasses manual model-building by learning an implicit policy that maps sensory delays and external forces directly to stable joint movements.
+-----------------------------------------------------------------------+
| Learning Framework |
| |
| +-----------------------+ +--------------------------+ |
| | Physical Priors | | Data-Driven Exploration | |
| | (Hamiltonian/Symplectic) | (RL / Meta-Learning) | |
| +-----------+-----------+ +------------+-------------+ |
| | | |
| +-------------------+-------------------+ |
| | |
| v |
| +------------------------------+ |
| | Coordinated Control Policy | |
| | for Underactuated Dynamics | |
| +------------------------------+ |
+-----------------------------------------------------------------------+
However, learning-based control for underactuated systems introduces its own hurdles. As Sinigaglia et al. point out, designing reward functions for these systems is notoriously difficult; because the agent lacks full control authority, providing dense feedback can easily introduce unintended behavioral biases. To mitigate this, researchers often combine DRL with Evolutionary Strategies to optimize black-box metrics, or embed structural physical constraints directly into the neural network architecture (such as using Symplectic ODE-Nets to learn underlying Hamiltonian dynamics).
Furthermore, underactuated systems are highly sensitive to external disturbances because they cannot directly cancel forces in unactuated directions. To address this, Richards et al. introduced an adaptive meta-learning framework. Instead of training a neural network to perfectly fit historical trajectory data, their framework optimizes for downstream tracking performance in closed-loop simulations. This ensures that when an underactuated platform like a quadrotor encounters unpredictable wind fields, the learned features map directly to effective, physically realizable compensation strategies.
Key Foundational Work
Detumbling and partial attitude stabilization of a rigid spacecraft under actuator failure (2000) — contributed 4 concept elements to this direction.
Characterization of Local Configuration Controllability for a Class of Mechanical Systems (2000) — contributed 3 concept elements to this direction.