Senxe Cerebellum: Biologically-Grounded Robotic Motor Control

Senxe Cerebellum is an open-source research framework that interfaces living biological neural networks (via the Cortical Labs CL1 microelectrode array platform) with high-precision industrial robotic manipulators.

The framework maps multi-modal physical sensor readings (force, torque, kinematics) into closed-loop electrical stimulation patterns and decodes biological firing outputs (spikes) into continuous action trajectories to solve force-sensitive assembly tasks (such as the RoboSuite NutAssembly benchmark).

Senxe Cerebellum v4.0 Live HUD Telemetry Overlay (RoboSuite NutAssembly task on Franka Panda)

Note

Hardware Fallback: This framework is built directly on the official Cortical Labs cl-sdk. It automatically detects physical hardware; when a CL1 device is not present, it gracefully falls back to the SDK's official Poisson simulation server, enabling developers and researchers to run the entire pipeline locally.

---

Core Scientific Modules

1. Neuromorphic Event-Driven Sparse Coding (VIE)

The Virtual Interference Encoding (VIE) module (core/vie.py) maps continuous environment observations onto the 64-channel microelectrode array (MEA).

• Sparse Encoding: To minimize signal crosstalk and cellular overstimulation, sensory modalities (e.g., force/torque deltas) are encoded using delta-tracking. Electrodes are only stimulated when physical quantities change significantly.
• Attention Multiplexing: Channels are dynamically reallocated between visual/tactile sensory modalities based on the robot's current task phase (searching vs. inserting).

2. Antagonistic Muscle-Pair Decoding

Motor outputs are decoded based on the biological flexor/extensor antagonistic principle (core/decoder.py).

• Opposing Populations: The 64 channels are interleaved into opposing sub-populations (Even/Odd pairs). Each of the 7 action dimensions is driven by the differential activity: $$\text{Action}[i] = \frac{\text{flexor} - \text{extensor}}{\text{flexor} + \text{extensor} + \epsilon}$$
• EMA Inertia Filter: Outputs are smoothed using an Exponential Moving Average (EMA) filter to mimic the biomechanical damping and inertia of physical muscle tissue, producing jerk-free trajectories.

3. FEP-Driven Kinematic Gate (PDI)

Rather than using hand-tuned exploration schedules, exploration is regulated by the Physical Disturbance Index (PDI) (core/pdi.py). Inspired by the Free Energy Principle (FEP):

• High PDI (unstable kinematics, high sensory surprise) increases Gaussian perturbation to force exploration and surprise minimization.
• Low PDI (stable kinematics, low surprise) limits perturbation to exploit the current steady-state control policy.

4. Intrinsic Firing-Rate Curiosity

The Neural Curiosity module (core/curiosity.py) monitors firing pattern novelty. Novel electrophysiological patterns boost the exploration rate, driving the neural network to escape local minima in silent or repetitive states.

---

System Architecture

graph TB
    subgraph Env [RoboSuite Environment]
        RS[Franka Panda Arm] -->|Tactile Sensory| F["Force (3D) & Torque (3D)"]
        RS -->|Kinematics| V["EEF Position & Velocity"]
        RS -->|Spatial Target| T["Peg-to-Hole Target Vector"]
    end

    subgraph Enc [Virtual Interference Encoding]
        F -->|Sparse Delta Coding| SDC["Delta Change Filter"]
        V -->|Attention Multiplexing| AMUX["Task Stage Router"]
        T -->|Tuning Curves| TC["Spatial Tuning (27 bins)"]
        SDC & AMUX & TC -->|Homeostatic Gain| Gain["Gain Adaptation (channel_gain)"]
        Gain -->|Pulse Design| Stim["64-ch MEA Stimulation Design"]
    end

    subgraph Bio [CL1 Wetware Platform]
        Stim -->|Electrical Pulse| MEA["64-ch Electrode Array"]
        MEA -->|Evoke Activity| Neu["Biological Neurons (STDP)"]
        Neu -->|Extracellular Recording| Spikes["Spike Train Extraction (threshold 99.5%)"]
    end

    subgraph Dec [Antagonistic Motor Decoder]
        Spikes -->|Even Channels| Flex["Flexor Population Sum"]
        Spikes -->|Odd Channels| Ext["Extensor Population Sum"]
        Flex & Ext -->|Differential Actuation| Diff["(Flex - Ext) / (Flex + Ext)"]
        Diff -->|Mechanical Inertia| EMA["EMA Smoothing Filter"]
        EMA -->|Continuous Control| Act["7D Joint Action Output"]
    end

    subgraph Gate [FEP Active Inference Gate]
        V -->|Velocity Variance| PDI["Physical Disturbance Index (PDI)"]
        Spikes -->|Firing Rate Novelty| Cur["Neural Curiosity"]
        PDI & Cur -->|Exploration Noise| Noise["explore_noise (Gaussian)"]
    end

    Act -->|env.step| RS
    Noise -->|Perturbation| Act
    RS -->|Reward & Collisions| FB["Stimulus Reinforcement"]
    FB -->|Predictable Stim| MEA
    FB -->|Unpredictable Noise| MEA

    %% Styles
    classDef envStyle fill:#0b132b,stroke:#48cae4,stroke-width:2px,color:#fff;
    classDef encStyle fill:#1c2541,stroke:#00b4d8,stroke-width:2px,color:#fff;
    classDef bioStyle fill:#102c57,stroke:#ff5757,stroke-width:2px,color:#fff;
    classDef decStyle fill:#1b4d3e,stroke:#2ecc71,stroke-width:2px,color:#fff;
    classDef gateStyle fill:#3a0ca3,stroke:#7209b7,stroke-width:2px,color:#fff;

    class RS,F,V,T envStyle;
    class SDC,AMUX,TC,Gain,Stim encStyle;
    class MEA,Neu,Spikes bioStyle;
    class Flex,Ext,Diff,EMA,Act decStyle;
    class PDI,Cur,Noise,FB gateStyle;

Loading

---

Major Updates (Compared to April 2026 Release)

Since the initial release (a1057ea on April 13, 2026), the framework has undergone major refactoring, bug fixing, and scientific alignment:

1. Critical Control Loop Fixes

• Double Action Scaling Bug: Resolved an issue where actions were scaled twice in both the Agent loop and the Antagonistic Decoder, which previously caused the robotic arm to stall.
• GymWrapper Flattening Fix: Bypassed GymWrapper observation flattening inside extract_obs. This restores access to structured observation dictionaries (native force, torque, and target vector values) from the MuJoCo simulation.
• Action Bias Normalization: Replaced an unconditioned, exponentially growing action_bias update with a bounded, clipped heuristic to prevent motor command divergence.

2. SDK Integration & Robustness

• Idempotent Context Management: Fixed a double-close bug in the cl_open() context manager that caused ClosedNodeError crashes in PyTables on exit. The Neurons.close() method is now monkeypatched to be fully idempotent.
• STDP Plasticity Sign Inversion: Corrected a biological STDP bug in the mock neuron simulator where Pre-before-Post spikes incorrectly triggered long-term depression (LTD) instead of long-term potentiation (LTP).
• NumPy 2.x Compatibility: Added shims to support running with NumPy 2.x, silencing internal deprecation warnings from the legacy parts of the cl-sdk.

3. Scientific Rigor & Benchmarking

• FEP Terminology Alignment: Deep-cleaned the codebase to replace reward-centric terminology (like "Dopamine Injection" and "Punishment") with information-theoretic terminology ("Predictable Stimulation" and "Unpredictable Stimulation"), aligning with the Free Energy Principle.
• Ablation Benchmark Suite: Added a dedicated benchmark runner (run_ablation_benchmark.py) and visualization utility (plot_ablations.py). It runs paired-seed trials to compare the biological agent against control groups (no-stim, zero-spikes, and randomized-spikes).
• Fair Baseline Comparison: Removed hindsight experience replay (HER) reward injection during the PPO evaluation loop to guarantee a scientifically honest comparison between biological and silicon baselines.

---

Quick Start

1. Installation

Install the necessary simulator and reinforcement learning baselines:

git clone https://github.com/AzurLiu/Senxe-Cerebellum.git
cd Senxe-Cerebellum
pip install -r requirements.txt
pip install cl-sdk

2. Configure MuJoCo Backend (macOS)

export MUJOCO_GL=glfw

3. Run the Biological Benchmark

Run the primary training script:

python senxe_demo_robosuite.py

This script runs the 7-DoF Franka Panda robot arm task, trains the biological agent, and saves a Cyberpunk-styled video overlay cl1_nutassembly.mp4 displaying the MEA grid, force telemetry, and live status watermarks.

4. Run the Ablation Study

To run the automated information-nullification benchmark:

python run_ablation_benchmark.py
python plot_ablations.py

This will run headless control trials and save a learning curve comparison to ablation_plot.png.

---

Author & License

• Author: Azur (Jiahao) — Independent developer, incoming University of Alberta student.
• License: Licensed under the MIT License (changed from AGPL v3 in June 2026).