Introduction: Beyond Collision Avoidance
Traditional robot navigation benchmarks often treat the environment as a static obstacle course where any contact is a failure. However, real-world indoor spaces are cluttered and dynamic. A robot operating in a home or office must be able to interact with its surroundings—pushing a chair to clear a path or opening a door to reach a room.
The Interactive Gibson Benchmark (iGibson 0.5) introduces "Interactive Navigation," a paradigm where robots are encouraged to interact with obstacles to optimize their paths. It provides a high-fidelity simulator that combines photorealistic rendering with physics-based interaction, alongside a new metric to balance path efficiency against the effort of disturbing the environment.
Technical Innovations
To support this new paradigm, the researchers introduced two major technical innovations to bridge the gap between photorealistic static scans and interactive synthetic environments.
1. Hybrid Rendering Engine
The original Gibson simulator relied on Image-Based Rendering (IBR), which was photorealistic but too slow for the massive iterations required by Reinforcement Learning (RL) and could not handle dynamic object movement. iGibson replaces this with a high-speed mesh renderer written in Python and C++. This engine supports real-time updates as objects are moved, achieving speeds up to 1,000 fps.
2. Multi-Modal Sensing and Annotation
The simulator provides a suite of sensor outputs necessary for complex navigation tasks, including RGB images, surface normals, semantic segmentation masks, and depth maps. To make static scans interactive, the authors developed a pipeline to "de-render" scenes. They used neural networks to identify objects, then had humans align high-quality CAD models to those locations, ensuring objects looked real but behaved according to physics using the Bullet engine.
The iGibson Dataset
The benchmark is built upon the Gibson V1 dataset, augmented to include:
- 106 indoor scenes featuring 1,984 interactable CAD models.
- High-Impact Categories: Focus on five categories: chairs, desks, doors, sofas, and tables.
- Realistic Clutter: Addition of ten common household items (baskets, shoes, toys) to each scene to create realistic obstacles.
- Robot Variety: Support for ten different robot models, including the TurtleBot v2 and the Fetch mobile manipulator.
Measuring Success: The Interactive Navigation Score (INS)
A key contribution of this work is the Interactive Navigation Score (INS). Unlike standard metrics that only measure path length, INS unifies path efficiency with effort efficiency. It penalizes the robot based on the mass of objects moved and the force applied.
Through semantic segmentation, robots learn that a cup is easy to move while a sofa is not. By adjusting an "interaction penalty" ($k_{int}$) in the reward function, researchers can train agents with different "personalities"—from cautious robots that avoid all contact to efficient agents that plow through light clutter to reach their goal faster.
Key Results and Insights
- Algorithm Comparison: Testing RL algorithms like PPO, DDPG, and Soft Actor-Critic (SAC) revealed that SAC provides the best balance between goal reaching and environmental preservation.
- Physical Realism: The simulator accurately reflects how a robot's own mass influences its strategy. A heavy Fetch robot naturally learns to push small obstacles that a lighter TurtleBot would have to navigate around.
- Generalization: Agents trained in iGibson demonstrate a strong ability to generalize to unseen floor plans, indicating that they learn fundamental concepts like object "push-ability."
Conclusion and Future Directions
iGibson 0.5 represents a significant shift toward socially and physically aware robotics. While current limitations include simplified physics parameters and a limited set of interactable object categories, the benchmark provides the foundation for robots that understand the world as a manipulatable space rather than a static map. Future versions aim to include more articulated objects (like cabinets and drawers) and a wider variety of household items to further close the gap between simulation and real-world domestic assistance.
Summary updated