Pandu (pandu-jev): Tiny, Local, Zero-API-Cost Policy Networks, Uncertainty Calibration, and Grounded Language Cortex
Pandu explores how minimalist local neural network policies (~2.9K–5K parameters) can serve as ultra-fast, zero-cost motor reflexes in closed-loop embodied environments, leveraging calibrated uncertainty for hybrid fallback and grounding semantic representations from frozen foundation language models via the Canonical Intent Protocol.
- Ultra-Compact Footprint:
- Pandu Core: 2,916 parameters (~11 KB) — zero external dependencies.
- Grounded Pandu Policy: 4,980 parameters (~19 KB) — grounds 16d environment state + 16d semantic intent.
- Microsecond Reflex Latency: 0.017 – 0.024 ms (17–24 microseconds) policy evaluation on standard CPU/Apple Silicon.
- Zero API Cost: Runs 100% locally with zero external API calls or token billing.
- Calibrated Uncertainty: Expected Calibration Error (ECE) of 0.86%; error detection AUROC of 0.952.
- Hybrid Fallback Advantage: Matches 100% Teacher-grade success while reducing average latency by 70.3% and monetary cost by 69.1%.
- Canonical Intent Protocol: Decouples cognitive language understanding from real-time control, ensuring complete invariance to linguistic variation across 5 OOD stress tiers.
pandu-jev/
├── bin/ # Executable CLI helper scripts
│ └── pandu-jev # Standalone CLI runner
├── src/ # Core library sources
│ ├── cli.py # Click/Rich interactive CLI interface
│ ├── env/ # Universal closed-loop environments
│ │ ├── base.py # Abstract universal environment interface
│ │ ├── gridworld.py # GridWorld with dynamic hazards and procedural maps
│ │ └── racing.py # 2D continuous vehicle dynamics environment
│ ├── models/ # Policy and adapter architectures
│ │ ├── policy.py # Canonical TinyPolicy (2.9K) and scalable MLP tiers
│ │ └── grounded.py # GroundedPanduPolicy (4.9K) & CanonicalIntentProtocol
│ ├── expert/ # Demonstration oracles
│ │ └── astar.py # Optimal A* pathfinding oracle
│ ├── arena/ # Multi-agent arena & interactive game GUIs
│ │ ├── base.py # Arena framework, bot protocol & tournament engine
│ │ ├── snake.py # Competitive 2-player snake & anti-collision bots
│ │ ├── snake_gui.py # Multi-level Snake HTTP server & game engine
│ │ ├── chess_env.py # Micro chess environment, features & heuristics
│ │ ├── chess_gui.py # Interactive Chess GUI server & move history engine
│ │ ├── web_snake/ # Zero-dependency HTML5 Canvas Snake interface
│ │ └── web_chess/ # Zero-dependency HTML5 Canvas Chess interface
│ ├── datasets/ # Demonstration collection & serialization
│ │ ├── trajectory.py # Core spatial expert trajectories
│ │ └── language.py # 5-tier OOD language-grounded trajectories & Oracle Intent
│ ├── training/ # Training pipelines
│ │ ├── bc.py # Supervised Behavioral Cloning + temperature scaling
│ │ └── ppo.py # Reinforcement Learning (PPO + GAE actor-critic)
│ └── evaluation/ # Calibration & uncertainty metrics
│ ├── calibration.py # ECE, MCE, Brier score, reliability diagrams
│ ├── uncertainty.py # OOD stress regimes (unseen, blocked, noisy)
│ └── stress_testing.py # Environmental robustness suite (Versions A–E)
├── experiments/ # Empirical research benchmark suites
│ ├── benchmarks/ # Scaling & hybrid runtime benchmarks
│ │ ├── model_scaling.py # Parameter scaling laws (3K, 50K, 1M, 5M)
│ │ └── hybrid_fallback.py # Teacher vs Pandu vs Hybrid fallback runtime
│ ├── language/ # Language grounding & HF model tests
│ │ ├── grounded_cortex.py # 5-tier OOD grounded language benchmark
│ │ ├── modernbert.py # Standalone ModernBERT-Tiny HF benchmark
│ │ ├── smollm2.py # Standalone SmolLM2-135M HF benchmark
│ │ └── qwen3.py # Standalone Qwen3-0.6B HF benchmark
│ ├── results/ # Benchmark output artifacts & JSON logs
│ ├── run_all_tests.py # Master test runner (all suites)
│ └── run_research.py # Master empirical research pipeline
├── docs/ # Research specifications & empirical papers
│ ├── PLAN.md # Research and prototype plan
│ ├── RESEARCH.md # Publication-grade empirical research report
│ └── GROUNDED_LANGUAGE_ARCHITECTURE.md # Grounded Language Cortex specification
├── tests/ # Modular Pytest suite
│ ├── test_arena.py # Chess, Snake, GUI engines & tournament tests
│ ├── test_env.py # GridWorld and continuous racing dynamics tests
│ ├── test_expert.py # A* pathfinding oracle tests
│ ├── test_models.py # TinyPolicy (2.9K) architecture and scaling tests
│ ├── test_training.py # Behavioral cloning & trajectory dataset tests
│ ├── test_evaluation.py # Calibration, ECE & Brier score tests
│ └── test_grounded_language.py # Grounded policy, adapter & protocol tests
└── pyproject.toml # Packaging configuration
# Clone the repository
git clone https://github.com/Haslab-dev/pandu-jev.git
cd pandu-jev
# Create virtual environment and install editable package
python -m venv .venv
source .venv/bin/activate
pip install -e .Experience the local policy navigating with live calibrated confidence and step visualization:
pandu-jev play gridworld(or ./bin/pandu-jev play gridworld)
Pandu includes zero-dependency browser-based GUIs powered by Python's native http.server, HTML5 Canvas, and WebAudio API.
Features auto-run turn-based gameplay, interactive move history scrubber with FEN reconstruction, live sub-millisecond bot decision latency telemetry, material advantage counters, and robust resolution for checkmates, 3-fold repetitions, and insufficient material draws.
pandu-jev arena chess-gui
Features Easy, Medium, and Hard difficulty levels, Solo, Vs AI, and AI-vs-AI spectator modes, toroidal wall-wrap mode (default in Easy mode or toggleable), protected spawn runways, anti-suicide hazard filtering, and an asynchronous input buffer queue.
pandu-jev arena snake-gui --level easy
# Run quick benchmark suite
pandu-jev benchmark --quick
# Run Grounded Language Cortex & Canonical Intent Protocol benchmark
pandu-jev test-language
# Run master empirical research suite (Phases 3-14)
python experiments/run_research.py# Run all 25 unit and integration tests via PyTest
pytest tests/ -v
# Run unified test runner (Unit tests + HF models + Language cortex)
pandu-jev test-all KNOWLEDGE (Cognitive Cortex)
│
Foundation LM (Frozen)
SmolLM2 / ModernBERT / Qwen3
│
semantics
│
▼
┌─────────────────────────┐
│ Pandu Intent Protocol │ (Canonical Semantic Boundary)
└────────────┬────────────┘
│
▼
CONTROL (Motor Reflex Policy)
┌─────────────────────────┐
│ Pandu │ (Trainable ~4,980 params)
└────────────┬────────────┘
│
fast action (<0.03 ms)
│
▼
WORLD
"The LLM thinks about what is meant. Pandu decides what to do right now."
- Comprehensive Empirical Research Paper: docs/RESEARCH.md
- Grounded Language Cortex Specification: docs/GROUNDED_LANGUAGE_ARCHITECTURE.md
- Development & Prototype Plan: docs/PLAN.md