To control a complex physical body—whether it is a biological organism with hundreds of muscles or an anthropomorphic robot with dozens of actuators—a control system faces the "degrees of freedom" problem. If every joint or muscle is regulated independently, the control space explodes exponentially. This redundancy introduces severe computational overhead, susceptibility to noise, and destructive crosstalk between overlapping commands. Yet, biological systems execute smooth, real-time movements effortlessly.
The insight driving this research direction is that high-dimensional motor systems do not actually operate in high dimensions. Instead, the central nervous system and advanced robotic controllers project complex physical dynamics onto a highly compressed, low-dimensional manifold. By grouping individual actuators into functional, synchronized units called motor synergies or primitives, the system transforms a mathematically intractable control problem into a low-dimensional optimization task.
Why Compression Works: The Geometry of Tasks
Why can a system discard 90% of its control dimensions without losing dexterity? The answer lies in the statistical structure of physical tasks. Movements are rarely random; they are constrained by environmental physics, skeletal geometry, and task goals.
Data-driven analyses of human movement consistently reveal this underlying simplicity. For instance, the authors of Human hand modelling: kinematics, dynamics, applications note that approximately 80–90% of the variance in human grasping can be captured by just two or three principal components. Similarly, in daily manual activities, the authors of Decoding of human hand actions to handle missing limbs in neuroprosthetics demonstrate that the hand's complex kinematic space effectively collapses into a subspace of just four to five dimensions. For whole-arm movements, Averta et al. show in Exploiting upper-limb functional principal components for human-like motion generation of anthropomorphic robots that three functional components explain over 80% of behavioral variance.
From an information-theoretic perspective, this compression is not just a biological limitation; it is an optimal strategy. Terence D. Sanger demonstrates in Optimal Unsupervised Motor Learning Predicts the Internal Representation of Barn Owl Head Movements that an optimal internal representation of motor control can be derived from the eigenvectors of the physical system's output statistics. By aligning internal control coordinates with these principal components, a system minimizes crosstalk and simplifies the mapping between high-level intent and low-level muscle activation. In a subsequent work, Terence D. Sanger frames motor control as a data compression task in Optimal Movement Primitives, showing that a remarkably small number of primitives can minimize reconstruction errors across diverse multi-dimensional movements.
Biological Implementation and Neural Common Drive
How does the nervous system physically execute this low-dimensional strategy? Rather than sending independent signals to every motor unit, the brain utilizes a "common drive" to activate groups of muscles simultaneously.
Using non-invasive neural interfaces, Silvia Muceli discovered in Non-invasive analysis of motor neurons controlling the intrinsic and extrinsic muscles of the hand that a single common neural signal accounts for over half of the activation variance across diverse muscle groups during complex hand tasks. This modularity is further supported by Borzelli et al. in Frequency-specific intermuscular coherence of synergistic muscles during an isometric force generation task, who show that the central nervous system regulates redundant musculoskeletal systems by organizing muscles into a small set of functional modules.
The functional necessity of these synergies becomes apparent when they are disrupted. Qi L. Xiong found in Inter-Limb Muscle Synergies and Kinematic Analysis of Hands-and-Knees Crawling in Typically Developing Infants and Infants With Developmental Delay that infants with developmental delays exhibit a reduced number of muscle synergies, leading to less smooth and more restricted movement patterns. This indicates that a healthy motor system relies on a rich yet structured library of low-dimensional primitives to compose complex behaviors.
Accelerating Robotic Learning and Control
For artificial systems, operating directly in high-dimensional action spaces makes reinforcement learning (RL) and motion planning computationally intractable. By constraining the search space to a low-dimensional synergy manifold, roboticists can drastically improve sample efficiency.
For example, Luck et al. demonstrate in From the Lab to the Desert: Fast Prototyping and Learning of Robot Locomotion that constraining a legged robot's control space to a small number of motor synergies rather than independent joint movements allows reinforcement learning algorithms to discover effective locomotion policies with significantly fewer physical trials. When training neural networks on complex musculoskeletal tasks, the authors of Acquiring musculoskeletal skills with curriculum-based reinforcement learning observed that artificial agents naturally evolve low-dimensional motor synergies mirroring biological systems. This emergence suggests that low-dimensional coordination is an inherent mathematical solution to managing redundant actuators in physical environments.
Furthermore, these synergies can emerge autonomously from generalized architectures. Gupta et al. demonstrate in MetaMorph: Learning Universal Controllers with Transformers that attention mechanisms within universal transformer controllers can automatically discover motor synergies across different robot morphologies, grouping multiple actuators into coordinated gaits without explicit manual engineering.
Reducing Cognitive Load in Human-Machine Interfaces
The principle of low-dimensional control is equally vital when a human is placed "in the loop," such as in teleoperation or neuroprosthetics. Forcing a human operator to manage dozens of degrees of freedom simultaneously imposes an overwhelming cognitive load.
To address this, shared-control frameworks delegate low-level coordination to the machine while leaving high-level task dimensions to the human. Lin et al. simplified the teleoperation of high-degree-of-freedom robotic hands in Learning Visuotactile Skills with Two Multifingered Hands by mapping multiple finger joints to a single collective "power grasp" command, allowing operators to execute complex bimanual tasks seamlessly. Similarly, Davies et al. demonstrate in Comparing Alternate Modes of Teleoperation for Constrained Tasks that narrowing an operator's responsibility to a low-dimensional task subset (such as 2D positioning) while letting the robot handle orientation significantly reduces mental fatigue and improves precision.
In prosthetics, mapping high-dimensional biological signals into low-dimensional bottlenecks provides smoother, more intuitive control. Vujaklija et al. utilized autoencoders in Online mapping of EMG signals into kinematics by autoencoding to compress high-dimensional electromyographic (EMG) signals into a low-dimensional representation of essential muscle synergies. This approach filters out channel noise and yields highly efficient, direct movement trajectories, proving that leveraging the low-dimensional nature of motor control is the key to bridging the gap between human intent and machine execution.
Key Foundational Work
Optimal Unsupervised Motor Learning Predicts the Internal Representation of Barn Owl Head Movements (1993) — contributed 3 concept elements to this direction.