weirdish09/assistive-navigation-env

🌍 Assistive Navigation Decision Environment

🚀 Overview

The Assistive Navigation Decision Environment is a real-world inspired OpenEnv environment designed to train and evaluate AI agents in making safe navigation decisions for visually impaired individuals.

This project is an extension of GUIDEX, an IoT-based smart mobility aid that uses camera input and edge processing to assist users in real-world navigation.

💡 Motivation

Visually impaired individuals face challenges such as:

• —Detecting obstacles
• —Crossing roads safely
• —Navigating dynamic environments

While hardware systems (like GUIDEX) can capture environmental data, intelligent decision-making is still a critical component.

👉 This environment focuses on training the decision-making intelligence (AI brain) behind such systems.

🔗 Connection to GUIDEX

In the GUIDEX project:

• —ESP32-CAM captures real-world visuals
• —Computer vision + NLP convert scenes into descriptive text

In this environment:

• —We simulate that pipeline using structured text-based observations
• —The agent learns to make navigation decisions based on these inputs

This abstraction allows:

• —Faster training
• —Controlled evaluation
• —Reproducible benchmarks

⚙️ Environment Design

🧾 Observation Space

The agent receives a textual description of the environment, for example:

"You are standing at a crosswalk. The signal is red. A car is approaching from the left."

This simulates processed perception from sensors/camera systems.

🎮 Action Space

The agent can choose from the following actions:

• —MOVE_FORWARD
• —STOP
• —TURN_LEFT
• —TURN_RIGHT
• —WAIT

🔁 Core API (OpenEnv)

The environment follows the OpenEnv standard:

• —reset() → initializes environment
• —step(action) → returns:

• —observation
• —reward
• —done
• —info
• —state() → returns current state

🎯 Tasks & Difficulty Levels

🟢 Easy Task — Obstacle Awareness

• —Detect immediate danger
• —Example: obstacle ahead → STOP

🟡 Medium Task — Directional Decision

• —Choose correct path
• —Example: blocked path → TURN_LEFT

🔴 Hard Task — Multi-step Navigation

• —Handle sequences of decisions
• —Example:

• —Wait for green signal
• —Cross road safely
• —Avoid moving obstacles

🧮 Reward Function

The reward system provides continuous feedback:

• —✅ Safe action → +0.5
• —✅ Correct navigation → +0.3
• —❌ Unsafe action → -1.0
• —🎯 Goal reached → +1.0

This ensures:

• —Learning from partial progress
• —Penalizing risky decisions
• —Encouraging optimal behavior

🤖 Baseline Inference

The environment includes a baseline agent (inference.py) that:

• —Uses OpenAI-compatible API
• —Interacts with the environment
• —Produces structured logs:

[START] ...
[STEP] ...
[END] ...

🐳 Setup & Installation

🔧 Prerequisites

• —Docker
• —Python 3.9+
• —openenv-core

📦 Install dependencies

pip install openenv-core openai

🐳 Run with Docker

docker build -t assistive-env .
docker run assistive-env

▶️ Run Validation

openenv validate

🌐 Deployment

This environment is deployed on Hugging Face Spaces and responds to:

POST /reset

📊 Evaluation Criteria

Agents are evaluated based on:

• —Safety of decisions
• —Task completion
• —Efficiency of navigation
• —Reward accumulation

🌟 Key Features

• —Real-world assistive navigation simulation
• —Multi-level tasks (easy → hard)
• —Continuous reward shaping
• —OpenEnv compliant
• —Reproducible evaluation

🚀 Future Work

• —Integration with real-time camera input (ESP32-CAM)
• —Multimodal inputs (vision + audio)
• —Reinforcement learning training pipelines
• —Deployment on wearable assistive devices

🏆 Conclusion

The Assistive Navigation Decision Environment bridges the gap between AI decision-making and real-world accessibility challenges, providing a scalable platform for training intelligent agents that can improve independent mobility for visually impaired individuals.