⚡ Core Innovations & Key Highlights
To achieve reliable autonomous navigation in physical humanoid robotics, an embodied agent must swiftly perceive complex environments, evaluate spatial risks, and output decisive locomotion commands without latency.
connectailab-rl-CHOCO-BRAIN-v1-link01 serves as an Embodied Spatial Reasoning Cortex for the LINK-01 Bipedal Humanoid Robot, purpose-built for real-time Dynamic Obstacle Avoidance across complex physical chokepoints. Powered by the JEV (Judgment-Evaluation-Verification) spatial protocol and calibrated via Deep Reinforcement Learning (PPO), this cognitive model ingests real-time 21-dimensional sensorimotor telemetry and outputs optimal discrete locomotion commands in 0.05 seconds with zero hallucination.
🎯 Architectural Highlights at a Glance
Table with columns: Core Dimension, Technical Specification, Engineering Significance| Core Dimension | Technical Specification | Engineering Significance |
|---|
| 🏢 Environmental Complexity | 5-Obstacle Slalom Maze (#8a790033) | Evaluated across an obstacle-dense 3.4m course with 5 clustered physical barriers |
| 📈 Trajectory Scale | 2,530 Total Questions (Train: 1,770) | Distilled from a high-volume dataset of 1,770 continuous real-world walking decisions |
| ⚡ Inference Latency | 0.05s (50ms Real-Time Locomotion Gating) | Converts 0.5s sensory windows into 5 discrete gait commands in 50ms with zero hallucination |
| 📡 Multi-Modal Telemetry | 21-Dimensional Spatial Observation | Ingests target vector, 11-ray Lidar proximity, 2D velocity, and 5 historical gait choices |
| 🧭 Continuous Navigation | 25.0s Average Course Traversal | Performs 1,770 continuous decisions over 25 seconds, stably converging to the target plant |
🏛️ Physical AI & Embodied Intelligence Lineage
This model marks the milestone release of Phase 3 (Embodied Decision Cortex) in Hwihwa Lab's autonomous physical intelligence lineage:
+------------------------------------------------------------------------+| Phase 1: Biological Connectome & Sensorimotor Reflex || [hwihwalab/neuro-robo-connectome] & [neuro-robo-studio] || - Drosophila Whole-Brain MaleCNS & FlyWire biological connectome || - 5.67ms low-latency locomotion, chemotaxis, and obstacle evasion |+-----------------------------------+------------------------------------+ | v+------------------------------------------------------------------------+| Phase 2: Executive Prefrontal Cortex & Business Reasoning || [hwihwalab/connectailab-rl-CHOCO-BRAIN_GUARD-v1] || - 6-Story CONNECT TYCOON enterprise simulation environment || - Natural language JEV adjudication & fraud risk mitigation |+-----------------------------------+------------------------------------+ | v+------------------------------------------------------------------------+| Phase 3: Embodied Decision Intelligence for Bipedal Robotics || [hwihwalab/connectailab-rl-CHOCO-BRAIN-v1-link01] (Current Model) || - Environmental Telemetry Discretization (Lidar Raycast + Kinematics) || - Policy Distillation (PPO Brain -> LoRA SFT -> RLM Policy Gradient) || - Real-time Locomotion Gating for LINK-01 Bipedal Humanoid |+------------------------------------------------------------------------+
🎯 The 4 Core Navigation Missions
Trained inside the CONNECT AI WORLD Simulator, the decision model masters 4 core robotic navigation benchmarks:
Table with columns: Mission, Objective, Environmental Dynamics, Sensory Focus| Mission | Objective | Environmental Dynamics | Sensory Focus |
|---|
| 1. Goal Reaching | Navigate directly toward the target mug | Open/Dynamic terrain | Heading vector & distance minimization |
| 2. Target Picking | Identify and approach a specific target among candidates (Lantern, Apple, Mug, Plant) | Multi-target semantic selection | Selective goal orientation & visual chemotaxis |
| 3. Obstacle Avoidance | Evade obstacles (Barrels, Wooden Boxes, Cardboard, Power Distribution Boxes) | Cluttered corridors & chokepoints | Frontal & oblique raycast proximity gradients |
| 4. Path Following | Accurately trace complex waypoints without drift | Curved/Arbitrary trajectories | Lateral deviation error & cross-track correction |
🔄 Dual-Cortex Embodied Architecture
+----------------------------------------------------------+| CONNECT AI WORLD / MuJoCo 3D Physical Environment || - Cluttered Courses: Barrels, Boxes, Distant Targets |+-----------------------------+----------------------------+ | Multi-objective PPO v+----------------------------------------------------------+| [Policy Brain] LINK-01 PPO Brain Core || - File: brain.json (Sensory-Motor Policy Matrix) || - Discretized Decision Interval: 0.5 seconds |+-----------------------------+----------------------------+ | Telemetry Textualization (0.5s windows) v+----------------------------------------------------------+| [Reasoning Cortex] CHOCO-BRAIN-v1-link01 || - Qwen3.5-0.8B LoRA Adapter (Natural Language JEV) || - Actions: Forward, Left, Right, Turn-L, Turn-R || - 0.05s Real-time Navigation Command Output |+-----------------------------+----------------------------+ | Low-level Gait Activation v+----------------------------------------------------------+| [Physical Actuation] LINK-01 Bipedal Locomotion CPG || - 2D Kinematic Approximation & MuJoCo Gait Physics |+----------------------------------------------------------+
📊 Benchmark & Evaluation Results
🧪 Evaluation Environment: High-Density Obstacle Maze (#8a790033)
- Dataset Scale: 2,530 Total Sensory Questions (
Train: 1,770 | Valid: 254 | Test: 506) · 2.5× to 3× larger trajectory volume than typical baseline benchmarks (500–700 questions)
- Course Topography: Long-distance navigation (3.4m linear span from Start
[-1.7, 0.33] to Target [1.7, -0.54]) heavily obstructed by a 5-obstacle zigzag chokepoint (Barrels, Wooden Crates, Power Distribution Boxes, and Storage Units)
- Physical Hardness: Even the reference PPO Oracle Teacher (
brain.json), trained over 40,000 decision steps, records a 50.0% collision rate, establishing this benchmark as an extreme real-world stress test.
- Evaluation Protocol: 20 standardized simulated test trials across seeds 91001–91020.
Table with columns: Model Architecture, Test Question Accuracy, 20-Run Success Rate, Collision Rate, Avg Navigation Time, Selection Status| Model Architecture | Test Question Accuracy | 20-Run Success Rate | Collision Rate | Avg Navigation Time | Selection Status |
|---|
Teacher Policy Brain (brain.json) | 100.0% | 100.0% | 50.0% | 22.5s | Policy Oracle (RL Benchmark) |
| Zero-Shot Base LLM (Untrained) | 32.9% | 0.0% | 10.0% | — (Timeout) | Baseline |
| Supervised Fine-Tuned (SFT LoRA) | 72.5% | 25.0% | 80.0% | 25.0s | Winning Deployed Model (★) |
| Reinforcement Tuned (RLM) | 41.1% | 15.0% | 75.0% | 26.2s | Exploratory Variant |
💡 Benchmark Complexity Note (Obstacle Maze vs. Flat Sprint):
While open-field target-reaching tasks (goto_target) without barriers typically clear within 10–12 seconds, Course #8a790033 is an obstacle-dense 3.4m slalom corridor obstructed by 5 large physical barriers (Barrels, Wooden Crates, Power Distribution Boxes). The recorded 25.0s traversal represents 1,770 continuous real-time sensory evaluations and collision-evasion decisions rather than simple linear sprint velocity.
Engineering Insight: Moving beyond simplified test tracks, CHOCO-BRAIN-v1-link01 was deliberately evaluated on a complex 25.0s obstacle maze. Under this high-density stress environment, the 0.8B LoRA adapter achieved an impressive 72.5% decision fidelity with a 25.0% autonomous clearance rate.
🧭 Designer's Note: The Philosophy of Obstacles & Resilience
“A frictionless world produces fragile intelligence; real embodiment is forged through collisions.”
Robotic benchmarks frequently evaluate agents within idealized, obstacle-free corridors where high success rates are readily demonstrated. Yet reality—much like human life—never offers a frictionless environment. Physical spaces are naturally filled with immovable obstacles, irregular clutter, and narrow chokepoints.
When designing Course #8a790033, we deliberately chose not to clear the path. We placed 5 massive physical obstacles directly across the bipedal robot's 3.4-meter trajectory:
- Collisions are not failures; they are the sensory boundaries of reality. Just as a child learning to walk stumbles hundreds of times against furniture to grasp the physical contours of the world, LINK-01’s collisions are badges of courageous exploration across an unforgiving physical maze.
- Resilience over Avoidance. Under heavy penalization, an agent often freezes or retreats in fear. But true embodied intelligence does not surrender. Over 25 long seconds and 1,770 sensory decisions, CHOCO-BRAIN-v1-link01 repeatedly collides, recovers its dynamic balance, and persists toward the distant green plant (the symbol of life and goal).
This model is a tribute to human resilience: proving that intelligence is not the absence of obstacles, but the courage to navigate through them.
🚀 Quick Start & Integration
1. High-Level Decision Engine (connect-tycoon)
from huggingface_hub import hf_hub_downloadfrom connect_tycoon.link01.brain import WorldBrain # Download brain policy matrix directly from Hugging Facebrain = WorldBrain.from_file(hf_hub_download("hwihwalab/connectailab-rl-CHOCO-BRAIN-v1-link01", "brain.json")) # Execute 0.5s decision step# obs: [target_dist, target_angle, ray_left, ray_center, ray_right, speed_x, speed_y]action_idx = brain.act(obs)print("LINK-01 Gait Action:", brain.actions[action_idx])
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_model_id = "Qwen/Qwen3.5-0.8B"adapter_id = "hwihwalab/connectailab-rl-CHOCO-BRAIN-v1-link01" tokenizer = AutoTokenizer.from_pretrained(adapter_id)base_model = AutoModelForCausalLM.from_pretrained( base_model_id, torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base_model, adapter_id)
📂 Repository Structure
📦 connectailab-rl-CHOCO-BRAIN-v1-link01├── 🧠 adapter_model.safetensors # Fine-tuned LoRA Decision Weights (21.7 MB)├── ⚙️ adapter_config.json # PEFT / LoRA Adapter Configuration├── 🔤 tokenizer.json # Qwen Fast Tokenizer Vocabulary (20 MB)├── 🔤 tokenizer_config.json # Fast Tokenizer Runtime Configuration├── 🎬 replay.mp4 # Official Simulation Video Preview (18.1 MB)├── 📜 README.md # Official English Model Card├── 📜 README_KR.md # Official Korean Technical Manual├── ⚖️ LICENSE # Apache-2.0 Open Source License├── 🧠 brain.json # PPO Sensory-Motor Policy Matrix (122 kB)├── 🌍 world.json # 5-Obstacle Slalom Maze Environment (#8a790033)├── 🏷️ connectailab.json # Official Competition Verification Metadata├── 📊 eval_evidence.json # Raw 26.2s Traversal Benchmark Metrics├── 🔬 course_fingerprint.json # Spatial Lidar & Course Complexity Analysis└── 📂 assets/ # High-Resolution Benchmark Dashboard Charts
🛠️ Intended Uses, Limitations & Sim-to-Real Scope
1. Primary Intended Uses
- Physical Robotics Research: Real-time obstacle avoidance and autonomous navigation policy distillation for bipedal humanoid architectures.
- Embodied Cognition: Multimodal spatial telemetry reasoning via JEV (Judgment-Evaluation-Verification) protocol, bridging language model reasoning with physical actuation.
- Reinforcement & Imitation Learning: Benchmark reference for policy distillation from PPO teacher models into fine-tuned LoRA causal language models.
2. Known Limitations & Domain Boundary
- Kinematic Approximation Scope: The evaluation harness operates on a 2D kinematic gait approximation calibrated from MuJoCo physics. When deploying to real LINK-01 hardware, local joint torque limits, ground friction coefficients, and mechanical backlash must be addressed via Sim-to-Real domain randomization.
- Decision Granularity (0.5s): This cognitive model serves as an executive cortex operating at 2 Hz (0.5s per step). High-frequency dynamic balance stabilization (100 Hz+) remains governed by the embedded low-level CPG motor controller.
- Static Clutter Specialization: Optimized for complex static physical obstacles (barrels, boxes, cabinets). Dynamic obstacle tracking (moving pedestrians, vehicles) represents a planned Phase 4 milestone.
- Domain Specialization: Tailored strictly for spatial robotic navigation, obstacle avoidance, and target retrieval. It does not perform open-domain conversational chat.
👤 About Hwihwa Lab
“AI × Automation × Aesthetics — Crafting tools you'll love ✨”
Hwihwa Lab is an independent AI research & engineering studio dedicated to blending practical on-device intelligence, autonomous operations, and refined aesthetics. We engineer resilient systems that free solopreneurs and indie creators from operational overhead—giving people their time and creative joy back.
🔬 Core Research & Engineering Pillars
- Zero-Cost Edge Operations: Private, on-device AI pipelines for autonomous business workflows without ongoing cloud API overhead.
- Autonomous Decision OS & Micro-ERP: Deterministic, prefrontal-cortex-style reasoning architectures designed to orchestrate daily business decisions and operations.
- Embodied AI & Domain-Specific Robotics: Simulation-driven reinforcement learning, physical control brains, and specialized domain data engineering for real-world environments.
- Interactive Creative Aesthetics: Fusing functional automation with refined visual craft and modern design philosophy to build delightful tools.
🌐 Connect & Research Ecosystem
Part of the CONNECT AI LAB Collaborative Research Network · LINK-01 Robotics Foundation
📄 License
This project, weights, and documentation are distributed under the Apache 2.0 License — see the repository LICENSE file for full details.
Base model weights adhere to the original Qwen License.
🏛️ 모델 정체성 (Identity)
connectailab-rl-CHOCO-BRAIN-v1-link01은 휘화랩(Hwihwa Lab)의 Physical AI 3부작(생체 커넥톰 ➔ 경영 전두엽 ➔ 신체화 대뇌) 중 제3단계에 해당하는 이족보행 휴머노이드 로봇(LINK-01) 전용 실시간 의사결정 대뇌입니다.
⚡ 핵심 엔지니어링 실측 성과 (At a Glance)
- 고난이도 실전 슬라롬 지형 실측 (
#8a790033): 단순 단거리 트랙이 아닌, 5대 대형 복합 장애물(드럼통, 배전함, 나무상자, 종이상자, 캐비닛)이 지그재그로 가로막힌 3.4m 장거리 미로 환경에서 정밀 검증
- 1,770개 대규모 궤적 증류: 2,530개 텔레메트리 문제 중 1,770개의 실전 보행 궤적을 Qwen3.5-0.8B LoRA 어댑터로 정밀 모방 학습
- 0.05초(50ms) 초저지연 게이팅: 21차원 다중 감각(라이다 11-Ray, 속도, 이전 선택) 데이터를 50ms 만에 환각(Hallucination) 없이 5대 보행 모터 명령으로 확정 출력
- 72.5% 의사결정 모사율: 극한의 물리 환경 속에서도 PPO 교사 두뇌의 판단을 72.5% 일치도로 재현하며 25초 연속 주파 달성
🧭 설계 철학 (Philosophy)
마찰이 없는 단순한 온실 지능을 지양하고, 수많은 장애물과 충돌에 직면하면서도 끝끝내 목표(화분)를 향해 전진하는 인간적 회복탄력성(Resilience)을 물리적 로봇 두뇌에 실현했습니다.
📘 자세한 2계층 신경망 구조, 2D 지형도 분석 및 빠른 시작 코드는 공식 한국어 매뉴얼인 README_KR.md에서 확인하실 수 있습니다.
Built and maintained by HWIHWA LAB in collaboration with CONNECT AI LAB.