Sparse Cosine Optimized Policy Evolution (SCOPE)

• SCOPE is a method for evolutionary policy learning that applies a 2D DCT-II to compress high-dimensional pose-time matrices for hexapod gait generation.
• It reduces the controller’s input from 2700 scalars to 54 coefficients, drastically minimizing the search space and improving sample efficiency.
• The approach yields a 20% performance improvement over raw-input methods using a steady-state genetic algorithm, demonstrating effective search reparameterization.

Sparse Cosine Optimized Policy Evolution (SCOPE) is a method for evolutionary policy learning that addresses a central bottleneck in hexapod gait generation: as controllers ingest richer temporal and spatial observations, the number of free parameters required to map inputs to actions grows rapidly, inflating the search space and degrading convergence. In the reported hexapod setting, SCOPE moves policy learning into the frequency domain by applying a 2D Discrete Cosine Transform (DCT-II) to a pose-time matrix, truncating the coefficient matrix to a compact low-frequency block, and evolving the controller on those compressed coefficients rather than on raw inputs. On the Webots Mantis hexapod, this reduces the controller input from 2700 scalars to 54 and yields a statistically significant improvement in efficacy relative to a raw-input steady-state genetic algorithm baseline (O'Connor et al., 17 Jul 2025).

1. Problem formulation and central idea

SCOPE targets evolutionary policy learning in regimes where the input space contains long time-series and multiple coupled sensor or actuator channels. In such settings, a direct increase in input complexity causes a proportional increase in policy parameters, which in turn produces large chromosomes, slows search, and reduces sample efficiency. In hexapod gait generation, the controller must interpret high-dimensional time-series of robot poses and map them to gait parameters for 18 motors.

The method’s core idea is to compress the controller input before evolution, but to do so in a way that preserves structured temporal and spatial information. Rather than selecting raw features or learning a data-adaptive latent space, SCOPE applies the DCT-II to the input matrix and retains a top-left low-frequency block of coefficients. This exploits the DCT’s energy compaction property: dominant signal structure tends to concentrate in low frequencies, so a fixed-size block can preserve most of the signal energy while discarding fine-grained high-frequency content.

In the reported hexapod example, the original controller input is a matrix $X \in \mathbb{R}^{6 \times 450}$ formed from recent pose history. SCOPE truncates the resulting DCT coefficient matrix to a $6 \times 9$ block, flattening it to 54 inputs. The compression ratio is therefore

$r = \frac{k_1 k_2}{N M},$

which in the experiment is $54/2700 \approx 0.02$, corresponding to an approximately 98% decrease in input size.

2. Frequency-domain representation and truncation

SCOPE uses a 2D DCT-II. For an input matrix $X \in \mathbb{R}^{N \times M}$, the transform is

$$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$

for $u=0,\ldots,N-1$ and $v=0,\ldots,M-1$, with $\alpha(0)=\sqrt{1/N}$, $\alpha(u>0)=\sqrt{2/N}$, and analogously for the column dimension. By separability,

$6 \times 9$0

where $6 \times 9$1 and $6 \times 9$2 are the 1D DCT-II basis matrices.

The energy of the coefficient matrix is defined as

$6 \times 9$3

SCOPE does not select coefficients by global magnitude ranking. Instead, it takes the top-left low-frequency block

$6 \times 9$4

which preserves smooth spatial-temporal structure and avoids discontinuous selection across frequency bands.

The truncation is constrained by

$6 \times 9$5

The selected block is then flattened in column-major order into the controller input vector

$6 \times 9$6

A distinctive property of the method is that it supports arbitrary output shapes $6 \times 9$7 subject to those dimension-wise inequalities. In the hexapod study, $6 \times 9$8 and $6 \times 9$9, and the selected shape is $r = \frac{k_1 k_2}{N M},$0. Choosing $r = \frac{k_1 k_2}{N M},$1 preserves all inter-leg low-frequency content while truncating temporal frequencies to $r = \frac{k_1 k_2}{N M},$2.

3. Hexapod gait-generation pipeline

The controller operates on a pose-history representation. Each pose is a $r = \frac{k_1 k_2}{N M},$3 matrix $r = \frac{k_1 k_2}{N M},$4, with one row per leg. A row contains position, velocity, and acceleration for the coxa, femur, and tibia joints. Concatenating the most recent 50 pose matrices horizontally yields

$r = \frac{k_1 k_2}{N M},$5

that is, 2700 scalar inputs.

The online pipeline has five stages.

Input construction: at each controller update, the last 50 state matrices are concatenated to form $r = \frac{k_1 k_2}{N M},$6.

DCT and truncation: the system computes

$r = \frac{k_1 k_2}{N M},$7

extracts

$r = \frac{k_1 k_2}{N M},$8

and vectorizes it to

$r = \frac{k_1 k_2}{N M},$9

Evolved policy: a linear policy maps the compressed features to gait parameters,

$54/2700 \approx 0.02$0

The paper states that $54/2700 \approx 0.02$1 degenerates to a vector when a single output per input is used; in the reported implementation, the controller evolves a weight vector $54/2700 \approx 0.02$2 and bias vector $54/2700 \approx 0.02$3, giving

$54/2700 \approx 0.02$4

The output supplies 54 values total, corresponding to 3 values per motor for 18 motors: phase $54/2700 \approx 0.02$5, amplitude $54/2700 \approx 0.02$6, and offset $54/2700 \approx 0.02$7.

Motor actuation via CPG: each motor follows

$54/2700 \approx 0.02$8

Actuation is subject to smoothness constraints $54/2700 \approx 0.02$9 and joint angle limits: coxa from $X \in \mathbb{R}^{N \times M}$0 to $X \in \mathbb{R}^{N \times M}$1, femur from $X \in \mathbb{R}^{N \times M}$2 to $X \in \mathbb{R}^{N \times M}$3, and tibia from $X \in \mathbb{R}^{N \times M}$4 to $X \in \mathbb{R}^{N \times M}$5. Transitions are linearly sliced to avoid jumps.

Fitness evaluation: each episode lasts 15 seconds and is split into five 3-second sub-episodes. At the start of each sub-episode, $X \in \mathbb{R}^{N \times M}$6 is rebuilt, the DCT is recomputed, and a fresh $X \in \mathbb{R}^{N \times M}$7 is emitted. The episode fitness $X \in \mathbb{R}^{N \times M}$8 is the Euclidean distance traveled in 15 seconds. No auxiliary penalties are added; stability and energy behavior are handled implicitly through joint limits and $X \in \mathbb{R}^{N \times M}$9.

4. Evolutionary optimization and empirical findings

The evolutionary backbone is a Steady-State Genetic Algorithm (SSGA), used both as the baseline and as the host algorithm for SCOPE (O'Connor et al., 17 Jul 2025). The baseline evolves directly over raw inputs, whereas the SCOPE variant evolves over truncated DCT coefficients.

The reported configuration is fixed across methods: population size 100, tournament selection of size 5, top two selected for reproduction, single-point crossover producing two offspring, Gaussian mutation with rate 0.003 and $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$0, and 5000 generations per run. Offspring are evaluated sequentially in simulation, replacing the worst two in the selected tournament. The simulation continues without resets; controllers are swapped in real time at evaluation boundaries.

The chromosome-size reduction is substantial. The raw-input baseline uses 2700 weights and 2700 biases, for 5400 parameters. SCOPE reduces this to 54 weights and 54 biases, for 108 parameters.

The experiments use Webots simulation of the Mantis hexapod on flat terrain with standard physics. The robot has 6 legs and 18 motors, all under positional control. The evaluation metric, termed efficacy, is Euclidean distance traveled in 15 seconds. Each method is run for 500 independent runs, with 5000 generations per run.

Method	Average fitness	Parameters
SCOPE	14.242	108
Baseline	11.880	5400

SCOPE produces higher mean fitness throughout training. The average fitness over trials is 14.242 for SCOPE and 11.880 for the baseline, corresponding to an approximately 20% increase. A one-sided Mann–Whitney U test gives $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$1 and $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$2, indicating statistical significance. The paper reports mean curves, but variance and confidence intervals are not reported.

The paper illustrates alternative truncation sizes qualitatively, but does not report quantitative ablations over $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$3 or over top-$$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$4 magnitude selection.

5. Computational properties and relation to alternatives

Because the DCT-II is separable, the 2D transform can be computed by applying 1D DCTs along rows and columns. With FFT-based implementations, the complexity is

$$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$5

For the reported hexapod setting with $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$6 and $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$7, the paper characterizes this overhead as small relative to simulation and evolutionary evaluation time.

Truncation itself is simple submatrix extraction with cost $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$8. Memory for the policy input drops from 2700 scalars to 54. The transform is recomputed dynamically at each 3-second update using the latest 50 frames; no caching is required at the reported scale.

The paper positions SCOPE against several alternative representation strategies. Relative to PCA or SVD, the DCT is deterministic, fast, and does not require covariance estimation or retraining under nonstationary data, but it is not data-adaptive. Relative to wavelets, the DCT is simpler and often sufficient when low-frequency structure dominates, though wavelets can capture transient events better. Relative to learned encoders such as autoencoders or attention models, SCOPE avoids additional training overhead and hyperparameters, functioning as a lightweight plug-in for evolutionary search.

The authors also emphasize domain generality. The same compression principle can be applied to other high-dimensional time-series in robotics, locomotion, manipulation, or UAV control, and to spatial-temporal observations that can be organized into matrices or tensors. Extension to 3D DCT is identified as a natural route for tensor inputs.

6. Assumptions, limitations, and implications

The method assumes that aggressive low-frequency truncation does not discard essential control information. In the reported pose-history domain, this assumption is acceptable, but the paper notes that fast contact transitions or slip events may reside in high-frequency components. In more contact-rich or noisy domains, overly small $$C_{u,v}= \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{M-1} X_{x,y}\cos\Big(\frac{\pi(2x+1)u}{2N}\Big)\cos\Big(\frac{\pi(2y+1)v}{2M}\Big),$$9 can degrade control.

Truncation sensitivity is therefore central. The paper recommends choosing $u=0,\ldots,N-1$0 so that the retained-energy ratio

$u=0,\ldots,N-1$1

exceeds a threshold such as 90–95%. This criterion is easy to compute online and provides a principled way to set compression strength.

The study also assumes precise, noise-free inputs. Under that assumption, percentile sparsification of coefficients—zeroing the lowest-magnitude $u=0,\ldots,N-1$2 values—was considered unnecessary and potentially harmful. The paper suggests that in noisy visual or proprioceptive domains, sparsification may help, but it is not used in the reported experiments.

A further limitation is architectural: the current controller flattens $u=0,\ldots,N-1$3 before applying a linear policy. The paper notes that more structured models, such as small convolutions over $u=0,\ldots,N-1$4, could exploit frequency locality while preserving compactness. This suggests a route for extending the approach without abandoning the compressed frequency-domain representation.

Within its reported scope, SCOPE shows that strong input compression can improve, rather than merely preserve, evolutionary control performance when the original input is highly redundant. In the hexapod study, the main effect is not only dimensionality reduction but a reparameterization of the search problem into a smaller, higher-signal space in which the steady-state genetic algorithm converges more effectively.