Spinal CPG

Central Pattern Generator — Innate Rhythmic Locomotion

The Central Pattern Generator (CPG) produces the baseline rhythmic pattern for quadruped locomotion. It is the “innate gait” — the creature can walk before the SNN has learned anything. The SNN then learns to modulate and eventually partially replace the CPG through the competence gate.

Biological Basis

CPGs are neural circuits in the spinal cord that generate rhythmic motor patterns without sensory feedback or descending commands. In vertebrates, CPGs produce locomotion patterns that can be modulated by brainstem signals (speed, direction) and sensory feedback (terrain adaptation). Grillner (2003) demonstrated this in lamprey, where isolated spinal cord segments produce swimming patterns.

Phase-Coupled Oscillators

Each leg has an independent oscillator with phase coupling to produce quadruped gaits:

phase_i(t+1) = phase_i(t) + 2π × frequency × dt

Joint commands from phase:
  hip_cmd  = sin(phase) × hip_amplitude
  knee_cmd = sin(phase + knee_offset) × knee_amplitude
  abd_cmd  = sin(phase × 0.5) × abd_amplitude

Phase offsets for trot gait:
  FL: 0°    FR: 180°   (diagonal pair)
  RL: 180°  RR: 0°     (diagonal pair)

Competence Gate

The CPG fades as the SNN demonstrates competence. This prevents the SNN from fighting the CPG pattern:

motor = CPG × cpg_weight + SNN × (1 − cpg_weight)

cpg_weight starts at 90%, decreases as actor_competence rises.
Actor competence is measured by sustained forward velocity
without falls.

In the 150k ball run, CPG weight dropped from 90% to 43% while actor competence rose to 0.939 — the SNN successfully took over most motor control.

Proximity Brake (Issue #79c)

When the creature approaches a ball (distance < 1.0m), CPG amplitude is reduced linearly:

proximity_amp_scale = max(0.1, 0.3 + 0.7 × (ball_dist / 1.0))

At 1.0m: full amplitude (1.0)
At 0.3m: 30% amplitude (0.3)
At 0.0m: minimum (0.1)

This prevents the creature from overshooting the ball. Biology: animals decelerate when approaching a target (optic flow regulation).

Steering

Steering is implemented inside the CPG via asymmetric per-leg amplitude modulation — not as a post-hoc motor offset. This follows the reticulospinal model (Grillner 2003): brainstem neurons modulate left/right CPG half-centers to produce turning.

References

Grillner, S. (2003). The motor infrastructure: from ion channels to neuronal networks. Nature Reviews Neuroscience
Ijspeert, A.J. (2008). Central pattern generators for locomotion control in animals and robots. Neural Networks

API Reference

SpinalCPG(n_actuators, joints_per_leg=4, config=None)

Phase-coupled oscillator CPG for quadruped gait.

compute(dt, arousal, freq_scale, amp_scale, steering) → np.ndarray

One step of CPG commands for direct-joint robots (Go2). Per-joint: [abd, hip, knee, ankle] × 4 legs. Steering via asymmetric abduction.

freq_scale — BehaviorPlanner multiplier (1.0=walk, 1.4=trot)
amp_scale — amplitude (1.0=normal, 0.1=proximity brake)
steering — -1..+1, reticulospinal turning model

compute_tendon(dt, arousal, freq_scale, amp_scale) → np.ndarray

CPG for tendon-actuated robots (Bommel). Output: [yaw, lift, extend] × 4 legs.

get_cpg_weight() → float

Current CPG weight in motor blend. Fades from 0.8 → 0.2 over cpg_weight_fade_steps.

from_evolved(path, n_actuators, jpleg) → SpinalCPG

Load evolved CPG parameters from cpg_config.json.

SpinalCPGConfig (key parameters)

frequency:         1.0 Hz (slow walk)
phase_offsets:     [0.0, 0.5, 0.75, 0.25] (walk gait)
hip_amplitude:     0.35   knee_amplitude: 0.25
base_amplitude:    0.30   max_amplitude:  0.70
cpg_weight_start:  0.8    cpg_weight_end: 0.2