SocNavGym is a comprehensive social navigation gymnasium environment for evaluating social robot navigation algorithms with multiple entity types including robot agents, human pedestrians, and obstacles (see Technical Specifications section). As part of the development team, I focused on reinforcement learning aspects and integration with predictive models (see My Contribution section). This was the primary environment for testing Predictive World Models and validating Adaptive Prediction Horizons (see Research Impact section), with results published in UKCI 2023 and my PhD thesis.
Target
Primary Environment
For predictive models
Book
UKCI 2023
Published results
Check
Team Project
2020-2024
We can see SocNavGym in action: multiple entities interacting in social navigation scenarios. Blue circle = robot agent, colored circles = human pedestrians, gray rectangles = static obstacles (tables, walls), green markers = goals/destinations (see Technical Specifications section for entity details)
2. Research Motivation
Why This Was Needed
Social robot navigation requires algorithms that can handle complex interactions with humans and dynamic environments. Traditional navigation environments lacked the social dynamics necessary to evaluate algorithms designed for real-world social scenarios.
As part of the development team, I contributed to SocNavGym to provide a comprehensive testing ground for social navigation algorithms with multiple entity types (see Technical Specifications section), enabling researchers to:
Evaluate Social Navigation Algorithms - Test algorithms in realistic social scenarios
Benchmark Performance - Compare different approaches systematically
Validate Predictive Models - Test predictive world models in complex environments (see Research Impact section)
Study Social Dynamics - Understand how robots interact with humans and other entities
Research Design
As part of the development team, I focused on reinforcement learning aspects and integration with predictive models (see My Contribution section). SocNavGym became the primary environment for testing Predictive World Models (2StepAhead, MASPM) and validating Adaptive Prediction Horizons (see Research Impact section), enabling comprehensive evaluation of sequential learning approaches in complex social navigation scenarios.
Multi-agent scenarios with robot, humans, and obstacles
Comprehensive Features
Detailed visual rendering of all entities
Full physics simulation for realistic movement
Complete state representation including all entity information
Research Integration
Compatible with standard RL frameworks
Supports predictive model evaluation
Enables comprehensive benchmarking
Extensibility
Customizable scenarios and entity configurations
Flexible evaluation metrics
Easy integration with new algorithms
5. Results & Performance
Research Outcomes
SocNavGym has been used successfully for testing Predictive World Models (2StepAhead, MASPM) and validating Adaptive Prediction Horizons in complex social navigation scenarios with multiple entities (see Research Impact section for details). Results were published in UKCI 2023 and used in my PhD thesis.
Social Rules Engine - Built-in social navigation conventions
Full Visual Rendering - Detailed graphics and visualization
RL Framework Compatible - Works with standard frameworks
Extensible Design - Customizable scenarios and metrics
8. Installation & Usage
Installation
# Clone the repository
git clone https://github.com/gnns4hri/SocNavGym.git
cd SocNavGym
# Install dependencies
pip install-e.# or
pip install gymnasium numpy pygame
When to Use SocNavGym
Use SocNavGym when:
Need comprehensive social navigation features
Require detailed visual rendering
Final evaluation and validation
Full-featured environment testing
Research publication validation
Use LiteSocNavGym when:
Rapid development and testing
Quick iteration during prototyping
Fast algorithm benchmarking
Basic Usage
importgymnasiumasgymfromsocnavgymimportSocNavGym# Create environment
env=SocNavGym()# Reset environment
obs,info=env.reset()print(f"Observation: {obs}")print(f"Info: {info}")# Run episode
forstepinrange(1000):# Get action from your model
action=your_model.predict(obs)# Step environment
obs,reward,terminated,truncated,info=env.step(action)# Render (optional)
env.render()ifterminatedortruncated:obs,info=env.reset()env.close()
Integration with Predictive Models
# Example: Using with 2StepAhead model
fromsocnavgymimportSocNavGymfromyour_modelsimportTwoStepAheadenv=SocNavGym()model=TwoStepAhead()obs,info=env.reset()forstepinrange(1000):action=model.predict(obs)obs,reward,terminated,truncated,info=env.step(action)ifterminatedortruncated:obs,info=env.reset()
SocNavGym was the primary environment for testing Predictive World Models (2StepAhead, MASPM) and validating Adaptive Prediction Horizons, with results published in UKCI 2023 and used in my PhD thesis. The environment enabled comprehensive evaluation of sequential learning approaches in complex social navigation scenarios, establishing a foundation for applying predictive models to real-world social robot navigation.
SocNavGym was important for evaluating and validating Predictive World Models and Adaptive Prediction Horizons. The environment enabled:
Testing Predictive World Models (2StepAhead, MASPM) with comprehensive evaluation
Validating Adaptive Prediction Horizons systems in complex social scenarios
Benchmarking navigation algorithms and evaluating RL approaches
Systematic performance analysis and research validation
Validating model performance in realistic social navigation scenarios before real-world deployment
10. Experimental Setup
Research Context
SocNavGym was used as the primary environment for testing Predictive World Models (2StepAhead, MASPM) and validating Adaptive Prediction Horizons (see Research Impact section). The environment enabled comprehensive evaluation of sequential learning approaches in complex social navigation scenarios with multiple entities including robot agents, human pedestrians, and dynamic obstacles.
Evaluation Framework
As part of the development team, I developed the evaluation framework for:
Predictive Model Testing - Testing 2StepAhead and MASPM models
Adaptive Prediction Validation - Validating Adaptive Prediction Horizons systems
Performance Benchmarking - Comparing different navigation approaches
vs. FallingBallEnv: Complex social navigation vs. simple physics simulation
vs. Standard Gym Environments: Social dynamics vs. basic navigation
14. FAQ
Q: What is SocNavGym?
A: SocNavGym is a comprehensive social navigation gymnasium environment for evaluating social robot navigation algorithms. It provides full-featured social navigation scenarios with multiple entity types including robot agents, human pedestrians, static obstacles, and dynamic obstacles (see Technical Specifications section for details).
Q: How was SocNavGym used in your research?
A: SocNavGym was the primary environment for testing Predictive World Models (2StepAhead, MASPM) and validating Adaptive Prediction Horizons (see Research Impact section). Results were published in UKCI 2023 and used in my PhD thesis.
Q: What was your contribution to SocNavGym?
A: As part of the development team, I focused on reinforcement learning aspects, predictive model evaluation framework, performance benchmarking systems, and integration with proposed models (see My Contribution section).
Q: When should I use SocNavGym vs LiteSocNavGym?
A: Use SocNavGym for final evaluation, comprehensive testing, and when you need full visual rendering. Use LiteSocNavGym for rapid development, testing, and prototyping (see Comparison section).
Q: Is SocNavGym compatible with standard RL frameworks?
A: Yes, SocNavGym is compatible with standard reinforcement learning frameworks and works with Gymnasium.
If you use SocNavGym in your research, please cite:
BibTeX:
@inproceedings{oguzie2023predictive,title={Predictive World Models for Social Navigation},author={Oguzie, Goodluck},booktitle={UK Conference on Intelligent Systems (UKCI 2023)},year={2023},organization={Aston University}}
APA:
Oguzie, G. (2023). Predictive World Models for Social Navigation.
In UK Conference on Intelligent Systems (UKCI 2023). Aston University.
Resources & Links
Technologies: Python • Gymnasium • Reinforcement Learning • Social Navigation