How do rotational neural dynamics produce accurate and flexible movement?

A robot body makes the causal question testable

Motor cortex recordings show that population activity passes through an attractor-like preparatory state and then rotates during movement execution, and this rotation appears across tasks including non-periodic reaches. What the rotation contributes to the actual movement has stayed unclear, because the motor system has far more neurons than muscles and recordings cover only a small fraction of neurons. The paper's premise is that a physical system whose controller can be observed completely removes both obstacles.

The authors built what they call a robotic analog: a fully analogous neuromuscular system with distributed artificial muscles, multimodal sensors and an artificial neural network controller trained by reinforcement learning. The muscles are braided liquid crystal elastomer fibers with a conductive coating that lets resistance report length; inertial sensors report joint orientation; motion capture tracks the end effector. Because every unit of the controller is visible and the body is real, the link between latent neural dynamics and physical movement can be examined directly rather than inferred.

Preparatory and execution activity occupy separate subspaces

The controller's actor network contains 512 recurrent neurons. Without any pre-designed task structure, their population activity split into two phases across tasks, recapitulating the preparatory and execution separation seen in motor cortex. The angle between the first principal components of the two phases was close to 90 degrees, with that first component explaining 20.39 plus or minus 3.10 percent of variance, and the full set of principal angles between the subspaces was broadly distributed, mostly between 30 and 90 degrees.

A variance-based analysis added that more than half of execution variance lay in the null space of the preparatory subspace, and more than half of preparatory variance lay in the null space of the execution subspace, across different feedback combinations, while subspace overlap stayed markedly below 0.5. The authors describe the result as largely segregated but not strictly orthogonal manifolds. Population variance also dropped sharply at the transition between states, matching the attractor-like convergence that sets initial conditions for execution. A local linear approximation captured neural state evolution with R2 between 0.9 and 1.0, so the system operated in a locally linear regime even though the neural-to-behavior mapping is nonlinear.

Rotations drive corrections perpendicular to the reach

During execution, the 512 neurons showed spiral-like trajectories, and applying jPCA, a method that fits a skew-symmetric matrix to the temporal derivatives of population activity to expose latent rotation, revealed consistent rotational dynamics in single-point reaching and in cycling tasks, and across all controller variants with different feedback combinations. Quantitatively, the angle between the neural state and its derivative approached 90 degrees in both vision-driven and proprioception-driven models, and the fit of a low-dimensional rotational model exceeded that from surrogate datasets preserving the original covariance structure. The rotations appeared in the recurrent layers but not in the input linear layer or the muscle output signals.

The paper's central mechanistic claim is that these rotations generate oscillatory maneuvers orthogonal to the reaching direction, which adjust the trajectory online and compensate for errors while maintaining stability. The authors report that this is confirmed by primate neural data, which they took from a publicly available monkey dataset. This is the paper's answer to what rotation is for: not a byproduct of recurrent wiring or sensory delay, but a means of steering correction in a direction that does not fight the reach itself.

Accuracy survives muscle damage, and more channels ease the controller

The robot arm maintained accurate trajectories despite various unexpected artificial muscle damages when visual feedback was available, and the paper reports reaching errors for different damaged scenarios. Rotational dynamics were not a fragile add-on: they appeared across controllers with different feedback combinations and with 3, 6 and 12 control channels, and the angle between neural state and derivative stayed near 90 degrees across those channel counts.

The channel comparison carries a second result. Enriching feedback and increasing control channels reduced the computational burden on the neural controller while maintaining motor performance. The authors explain this as richer inputs and outputs disambiguating sensorimotor mappings and giving the controller more redundant choices when searching for solutions to state estimation and motor planning. They call the underlying pattern counterintuitive neural energy principles under sensor and motor redundancies. The practical reading is specific to this system: redundancy helped because it made the controller's search easier, not because more actuators automatically improve a robot.

Rotations, preparation and accuracy appear together

Comparing controllers at different training stages exposed a coordinated jump. Sparsity of neural representations declined sharply and stabilized near the 2000th training episode across models with different feedback combinations. Stable rotational dynamics emerged only after 2000 episodes, alongside a jump and then plateau in PCA-explained variance and rotation radius. At the same point the arm suddenly began reaching targets accurately, with slower improvement afterward, and the separation between preparatory and execution subspaces shifted toward 90 degrees. Sudden changes in neural energy and muscle actuation energy occurred at the same episode.

The authors call these Eureka moments and interpret them as the emergence of a stable control regime in a closed-loop system, arising from interaction between neural dynamics and body mechanics. They are explicit that the detailed neural mechanisms behind the reorganization remain to be investigated. The transition was observed across architectures, feedback combinations and random initializations, though the exact episode varied slightly; both recurrent and feedforward architectures showed it, but recurrent layers developed stronger rotational dynamics and higher performance.

What the analog supports and what it leaves open

The evidence for function comes from a trained artificial controller and a physical robot, not from manipulating biological cortex. The primate comparison is a confirmation against an existing dataset, not an experiment in which rotations were perturbed in an animal. The authors argue that the subspace separation emerged without pre-designed structure and persisted across real and simulated settings regardless of feedback delays, which they use to argue the dynamics serve a functional purpose rather than being driven by sensory delays. That argument rests on the model's behavior, and the paper itself notes that the mechanism of the learning transition needs further work.

The scope is also bounded by the task set: single-point reaching and cycling movements, with feedback combinations and 3, 6 and 12 control channels. The reported robustness is to artificial muscle damage under the tested scenarios, and the energy findings are comparisons across control strategies and feedback combinations within this platform. None of this establishes that biological motor cortex uses rotations for exactly the same corrective role, only that a system optimized for reaching develops rotations that do this job.

Redundancy and orthogonal dynamics as design levers

For readers weighing what to take from the paper, the most decision-relevant comparison is between controller configurations rather than between the robot and the brain. Adding feedback and control channels lowered the neural controller's computational burden while preserving performance, and the rotations that supported correction appeared across 3, 6 and 12 channels. The paper therefore points to two levers for embodied control design: exploit orthogonal latent dynamics for online correction, and treat sensor and motor redundancy as a way to make the controller's optimization easier.

The authors connect this to humanoid robots and dexterous hands that coordinate large numbers of actuators, and to rehabilitation robots and prosthetic devices, while noting that their framework currently covers online closed-loop control and that extending it to mechanisms such as memory consolidation and replay is a future direction. The priority the evidence supports is narrower than general biomimicry: when a reaching controller must stay accurate under damage, the paper's results favor checking whether its latent dynamics rotate near orthogonally to the movement and whether added channels reduce, rather than merely relocate, the control burden.

Sources used in this answer