| --- |
| language: |
| - en |
| license: mit |
| tags: |
| - robotics |
| - embodied-ai |
| - mujoco |
| - humanoid |
| - reinforcement-learning |
| - evolutionary-strategies |
| - pcei |
| - mamba |
| - bitnet |
| - fastkan |
| pipeline_tag: robotics |
| --- |
| |
| # PCEI-Robotics: Embodied 18-Meter Ninja Warrior Humanoid Controller |
|
|
| [](https://github.com/MtnMCG/PCEI) |
| [](https://github.com/MtnMCG/PCEI) |
| [](https://github.com/MtnMCG/PCEI) |
|
|
| > **Official Weights, Physics Models, and Simulation Environments for the PCEI (Physically Constrained Emergent Intelligence) Embodied Humanoid Agent conquering the 18-meter American Ninja Warrior obstacle course in real-time MuJoCo continuous physics at 40 FPS.** |
| > |
| > Master GitHub Repository: [https://github.com/MtnMCG/PCEI](https://github.com/MtnMCG/PCEI) |
|
|
| --- |
|
|
| ## 🏃 Autonomous 18-Meter Obstacle Traversal Demonstration |
|
|
| Inside an American Ninja Warrior MuJoCo humanoid simulation, without dense Transformer attention or sample-inefficient deep RL, the PCEI agent autonomously accelerates down the high-momentum Sprint Runway (Stage 0), bounds across the angled 45° Quad Steps (Stage 1), and touches down cleanly at **18.00 meters** on the Stage 2 Balance Beam platform: |
|
|
|  |
|
|
| *▶️ [Watch High-Definition MP4 Video](best_run_18m_ninja.mp4)* |
|
|
| --- |
|
|
| ## 📊 Run Telemetry & Benchmarks |
|
|
| | Metric / Dimension | Measured Performance | Architectural Context | |
| | :--- | :--- | :--- | |
| | **Course Traversal Distance** | **18.00 meters** (Stage 2 Touchdown) | Cleared Stage 0 (0–8 m) & Stage 1 Quad Steps (8–18 m) | |
| | **Trajectory Duration** | **284 execution frames (7.1 s)** | Real-time continuous physics evaluation at 40 FPS | |
| | **Peak Sprint Velocity ($v_x$)** | **4.06 m/s (14.6 km/h)** | Dynamic acceleration across runway into 45° angled platforms | |
| | **Total Environment Training Steps** | **17.4M steps** | Traversal achieved 3×–5× faster than standard MLP / RL baselines | |
| | **Simulated Training Episodes** | **213,000+ episodes** | Sampled across 16 parallel MuJoCo vector environments | |
| | **Evolutionary Generations** | **110 generations** | Darwinian Pioneer-Anchor Partitioning (70% Anchors : 30% Pioneers) | |
| | **BitNet Codon Point Mutations** | **143,646 mutations** ($\\mu = 0.0025$) | Discrete ternary synaptic adaptation without catastrophic forgetting | |
| | **Start-Line Balance Retention** | **100.0%** (0 fall rate at origin) | Pioneer-Anchor duality prevented start-line forgetting | |
| | **Observation & Motor Dim** | **572-dim Input → 21-dim Torque** | 3D solid-state LiDAR + proprioception driving 21 continuous joint actuators | |
| | **Neural Backbone Engine** | **Mamba SSM + FastKAN + BitNet** | Constant `O(1)` recurrent state space + cubic B-spline non-linear torque manifold | |
| |
| --- |
| |
| ## 🧠 Architectural Foundation |
| |
| Standard monolithic Transformer architectures fail at high-frequency continuous control due to quadratic latency scaling ($\\mathcal{O}(L^2)$) and unbounded memory buffers. The PCEI robotic controller unifies three core principles: |
| |
| 1. **Selective State-Space Recurrence (Mamba SSM)**: Linear $\\mathcal{O}(L)$ compute with constant $\\mathcal{O}(1)$ memory buffer, updating continuous proprioceptive and LiDAR representations at 40 FPS without latency spikes. |
| 2. **FastKAN Cubic B-Spline Torque Manifolds**: Non-linear spline basis functions (`FastKANMixer`) placed in residual branches enable smooth, biologically plausible motor activations that prevent actuator jerk and joint instability. |
| 3. **Discrete 1.58-Bit Ternary Synaptic Gating**: Weight quantization constrained to `{-1, 0, +1}`, providing homeostatic metabolic efficiency and biological zero-inhibition. |
| |
| --- |
| |
| ## 🛠️ Reproduction & Local Execution |
| |
| To inspect the agent or run local rollouts in MuJoCo: |
| |
| ```bash |
| # Clone the master PCEI repository |
| git clone https://github.com/MtnMCG/PCEI.git |
| cd PCEI/pcei_framework |
| |
| # Create virtual environment and install dependencies |
| python3 -m venv venv |
| source venv/bin/activate |
| pip install -r requirements.txt |
| pip install mujoco gymnasium |
| |
| # Run interactive MuJoCo humanoid evaluation |
| python3 embodied_pcei/src/ninja_env.py --checkpoint checkpoints/pcei_ninja_best.pt |
| ``` |
| |
| --- |
| |
| ## 📜 Full Scientific Documentation & Compendium |
| |
| Complete specifications, mathematical proofs, and ablation studies are available in the master repository: |
| - Master Codebase & Research Papers: [https://github.com/MtnMCG/PCEI](https://github.com/MtnMCG/PCEI) |
| - [Research Paper IX: FastKAN Spline Manifolds & Mamba Motor Control](https://github.com/MtnMCG/PCEI/blob/main/pcei_framework/docs/RESEARCH_PAPER_09_FASTKAN_SPLINE_MANIFOLDS_AND_MAMBA_MOTOR_CONTROL.md) |
| - [Research Paper XI: Darwinian Pioneer-Anchor Population Partitioning](https://github.com/MtnMCG/PCEI/blob/main/pcei_framework/docs/RESEARCH_PAPER_11_DARWINIAN_PIONEER_ANCHOR_POPULATION_PARTITIONING.md) |
| |
| --- |
| |
| ## 🖋️ Citation |
| |
| ```bibtex |
| @article{ramirez2026pcei, |
| title = {Physically Constrained Emergent Intelligence: Subverting Transformer Hyperscaling via Linear State Spaces, Ternary Synaptic Gating, and Latent Dreaming}, |
| author = {Ramirez, Emerson and Antigravity Assistant}, |
| journal = {PCEI Theoretical Research Compendium}, |
| year = {2026}, |
| url = {https://github.com/MtnMCG/PCEI} |
| } |
| ``` |
| |
