Interactive Gibson Benchmark (iGibson 0.5): A Benchmark for Interactive Navigation in Cluttered Environments (1910.14442v3)

Abstract: We present Interactive Gibson Benchmark, the first comprehensive benchmark for training and evaluating Interactive Navigation: robot navigation strategies where physical interaction with objects is allowed and even encouraged to accomplish a task. For example, the robot can move objects if needed in order to clear a path leading to the goal location. Our benchmark comprises two novel elements: 1) a new experimental setup, the Interactive Gibson Environment (iGibson 0.5), which simulates high fidelity visuals of indoor scenes, and high fidelity physical dynamics of the robot and common objects found in these scenes; 2) a set of Interactive Navigation metrics which allows one to study the interplay between navigation and physical interaction. We present and evaluate multiple learning-based baselines in Interactive Gibson, and provide insights into regimes of navigation with different trade-offs between navigation path efficiency and disturbance of surrounding objects. We make our benchmark publicly available(https://sites.google.com/view/interactivegibsonenv) and encourage researchers from all disciplines in robotics (e.g. planning, learning, control) to propose, evaluate, and compare their Interactive Navigation solutions in Interactive Gibson.

• The paper introduces iGibson 0.5, a new benchmark that combines realistic interactive simulation with a novel navigation evaluation metric.
• It enhances photorealistic indoor environments by enabling object interactions and efficient training for reinforcement learning agents.
• Baseline results with PPO, DDPG, and SAC show that balancing path efficiency with interaction effort is key for improving robotic navigation.

An Overview of the Interactive Gibson Benchmark for Interactive Navigation

The paper "Interactive Gibson Benchmark (iGibson 0.5): A Benchmark for Interactive Navigation in Cluttered Environments" focuses on advancing the field of interactive navigation for robotic systems. Interactive navigation is a complex task where an autonomous agent must physically interact with its environment—such as moving objects—while navigating to a goal. This paper introduces a comprehensive benchmark, iGibson 0.5, which consists of two pivotal components that aid in the training and evaluation of interactive navigation solutions: a simulation environment and a novel evaluation metric.

Interactive Gibson Environment (iGibson 0.5)

The iGibson 0.5 simulation environment extends the original Gibson environment by incorporating the capacity for physical interactions with movable objects within photorealistic 3D reconstructed indoor scenes. These enhancements allow the simulation of realistic interactions such as pushing obstacles and opening doors, which are intrinsic to interactive navigation tasks. An important aspect of this environment is its efficient rendering capabilities, which significantly reduce computational overhead and improve training speeds for reinforcement learning agents. iGibson 0.5 provides pre-processed environments where common household objects have been annotated and integrated as interactable CAD models, facilitating realistic training scenarios.

Interactive Navigation Score (INS)

To evaluate interactive navigation capabilities, the paper introduces the Interactive Navigation Score (INS). This metric effectively captures the trade-off between two critical dimensions of navigation: path efficiency and interaction effort. Path efficiency relates to the success of reaching a destination using the shortest path, while interaction effort measures the extent of physical interactions with the environment. INS combines these dimensions, allowing researchers to evaluate the performance of an agent in terms of its navigation strategy and interaction with the environment.

Baselines and Evaluation

The authors present a series of baseline evaluations using well-established reinforcement learning algorithms such as PPO, DDPG, and SAC across different robotic platforms (e.g., TurtleBot v2 and Fetch). These baselines demonstrate how varying interaction penalties in the reward functions lead to different navigation strategies and emphasize how efficiently the agents balance path execution with interaction effort. The paper illustrates that the SAC algorithm consistently yields the best performance across various interaction penalties, maintaining a favorable balance between path and interaction efforts.

Implications and Future Directions

The findings of this paper contribute meaningfully to the theoretical framing and practical application of interactive navigation in robotics. The development of the iGibson 0.5 environment provides a standardized platform for testing and comparing navigation strategies in cluttered, realistic environments. The interactive navigation score offers a unified evaluation metric that could guide future developments in algorithm design and benchmark evaluations.

In a broader context, the insights garnered from such an approach can improve the deployment of robots in domestic and industrial settings, where the ability to navigate and manipulate objects autonomously is critical. Moving forward, extending the variety of object interactions and exploring solutions beyond RL-based methods might provide further robustness and versatility to the frameworks developed within iGibson 0.5. Researchers are encouraged to utilize these resources to facilitate further advancements in interactive navigation tasks, potentially leading to robots that can perform complex tasks in unstructured human environments more effectively.