The Brain3D visualization in MH-FLOCKE’s rendered videos now shows the actual cerebellar architecture — not an abstract network graph, but the real populations with their correct sizes and live spike activity from training data.
Six Populations, One Brain
The cerebellar SNN in MH-FLOCKE is organized into six biologically inspired populations, each with a distinct computational role:
- Mossy Fibers (MF) — sensory input from proprioception, CPG phase, and IMU
- Granule Cells (GrC) — sparse expansion layer, the largest population
- Golgi Cells (GoC) — inhibitory feedback, regulating granule cell activity
- Purkinje Cells (PkC) — the main learning substrate, driven by climbing fiber error
- Deep Cerebellar Nuclei (DCN) — motor correction output
- Output (OUT) — final motor commands to actuators
The Brain3D visualization now renders each population at its correct size. For the Freenove Robot Dog: 48 MF, 106 GrC, 18 GoC, 24 PkC, 24 DCN, 12 OUT = 232 total. For the Unitree Go2: 304 MF, 4000 GrC, 200 GoC, 24 PkC, 24 DCN, 72 OUT = 4,624 total.
Data-Driven, Not Decorative
Every aspect of the visualization is driven by actual training data stored in the FLOG (training log). The population sizes come from the FLOG metadata — written by the training script from the live SNN topology. Spike activity in the rendered video comes from the FLOG’s spike data recorded at each training step.
The FLOG metadata now includes a population_sizes dictionary that captures the exact neuron count per population. Both the Freenove and Go2 renderers read this data and pass it to the Brain3D overlay.
Scalable Architecture
The same cerebellar architecture scales from 232 neurons (Freenove, Raspberry Pi) to 4,624 neurons (Go2, desktop GPU). The topology.py module computes population sizes proportionally: for small networks, the granule cell layer shrinks but the architecture is preserved. Like a mouse cerebellum versus an elephant cerebellum — same cell types, same connectivity, different scale.
This scaling is what makes sim-to-real transfer possible: the Freenove brain is structurally identical to the Go2 brain, just smaller. A brain trained on one platform could theoretically transfer to the other with topology adaptation.
The source code and rendering pipeline are available on GitHub.