iGibson: Interactive Gibson Benchmark for Real-World Navigation and Manipulation

Bokui Shen, Fei Xia, Chengshu Li, Roberto Martin-Martin, Linxi Fan, Guanzhi Wang, Yunfei Bai, De-An Huang, Dinesh Jayaraman, Yuke Zhu, Li Fei-Fei, Anima Anandkumar, Isaac Kamwa, Jiajun Wu, Biao Jiang, Shengjie Lin, Yunhai Feng, Silvio Savarese · Stanford, NVIDIA, BIGAI, UT Austin · 2021

Summary

iGibson is a photorealistic, physics-based indoor simulation platform with 15 interactive scenes, 500+ object categories, and GPU-accelerated rendering, enabling RL and VLA training for mobile manipulation and embodied AI.

Abstract Summary

iGibson is the standard simulation platform for indoor robot navigation and manipulation research. Built on PyBullet for physics and iGibson renderer for photorealistic RGB, the platform provides 15 fully interactive scanned scenes (apartments, offices, houses) with 500+ object categories that can be opened, closed, picked, and placed. Unlike synthetic environments like Habitat or AI2-THOR, iGibson scenes are reconstructed from real-world scans, preserving accurate geometry, lighting, and object arrangements. iGibson 2.0 adds interactive physics: doors have hinges, drawers slide on rails, and liquids pour with particle simulation. This enables training policies for complex long-horizon tasks like "make coffee" (fill kettle → boil water → pour into cup) or "tidy the living room" (pick up objects → navigate to storage → place). The observation space includes RGB, depth, semantic segmentation, and scene graphs, making it compatible with both pure-vision RL and VLA models. The benchmark suite includes 50+ tasks: point-goal navigation, object-goal navigation, social navigation (avoiding virtual humans), rearrangement, and manipulation. Evaluation is standardized with generalization splits: train on apartment A, test on apartment B with different furniture layouts. iGibson is the evaluation platform for BEHAVIOR-1K, the largest embodied AI benchmark to date. For practitioners, iGibson installs via `pip install igibson` and integrates with Gymnasium, ROS2, and Habitat APIs. The asset downloader fetches scenes on-demand (total ~50GB). The included ROS2 bridge lets you run SLAM and navigation stacks (e.g., FAST-LIO2 + Nav2) in simulation before real-world deployment.

Key Points

15 photorealistic scanned scenes with 500+ interactive objects.
Real-world accurate geometry, lighting, and object arrangements.
Interactive physics: doors, drawers, liquids, deformable objects.
50+ tasks: navigation, rearrangement, manipulation, social nav.
ROS2 bridge + Gymnasium API for rapid sim-to-real transfer.

Extended Physical States for Simulation of Everyday Household Tasks

To perform household tasks, an agent needs to change objects’ states beyond their poses. iGibson 2.0 extends objects with five additional states: temperature, $T$, wetness level, $w$, cleanliness level (dustiness level, $d$, or stain level, $s$), toggled state, TS , and sliced state, SS .

Logical Representation of Physical States

The new extended object states from iGibson 2.0 are sufficient to simulate a new set of household activities in indoor environments. However, there is a semantic gap between the extended states (e.g. temperature or wetness level) and the natural description of activities in a household setup (e.g. cooking apples).

To bridge this gap, we define a set of functions that map the extended object states to logical states for single objects and pairs of objects.

Logical Representation of Physical States — iGibson: Interactive Gibson Benchmark for Real-World Navigation and Manipulation — (b) Object wetness level: (From left to right) A sink –toggleable droplet source– is toggled on, and starts to create droplets (second from left). Droplets can be contained in a receptacle to be poured on other objects (third from left). An object that can change its wetness level gets in contact with the droplets and absorbs them, changing its appearance (most right). With the droplets system, iGibson v2.0 provides a simplified liquid simulation sufficient to perform common household activities.

Virtual Reality Interface

iGibson 2.0’s new functionalities enable modeling new household activities and generating multiple instances in more densely populated scenes. To facilitate research in these new, complex tasks, iGibson 2.0 includes a novel virtual reality (VR) interface compatible with major commercially available VR headset through OpenVR [ 38 ] .

One of the goals is to allow researchers to collect human demonstrations, and use them to develop new solutions via imitation (see Sec. 6 ).

Virtual Reality Interface — iGibson: Interactive Gibson Benchmark for Real-World Navigation and Manipulation — (a) Object cleanliness level: (From left to right) An object is initialized with dust particles that can be cleaned with a cloth; Another object is initialized with stains that require a wet tool (a scrubber) to be removed; Realistic behaviors and motion are required in iGibson v2.0 to change the cleanliness level of the objects.

Evaluation

In our evaluation, we test the new functionalities of iGibson 2.0 explained above and that sets it apart from other existing environments. First, we create a set of tasks that showcase the new extended states (see Fig. 5(a) ) as their modification is required to achieve the tasks.

In our experiments, we make use of our discriminative and generative logical engine to detect task completion and create multiple instances of each task for training.

Evaluation — iGibson: Interactive Gibson Benchmark for Real-World Navigation and Manipulation — (b) Object toggling state: An object with the toggled state, an oven, is initially off. When it is toggled on, its appearance changes indicating the transition.

Sources and Demonstrations

Figures are reproduced from Shen et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.