Habitat v0.1.4 released: rigid body physics, object interaction, DDPPO, VLN, Depth/RGB noise models

TL;DR -- building a foundation towards next EmbodiedAI challenges: objects interaction, physics, sim2real, while building a benchmark for navigation tasks.

Habitat team has released the latest stable version (v0.1.4) of Habitat-Sim and Habitat-API-Tasks. The post about the version release.

HABITAT NEW FEATURES HIGHLIGHTS:

• Rigid body physics support improvements: Visual and collision bounding boxes, Compound object support, Configurable object scaling, Discrete contact testing.
• A new large-scale distributed reinforcement learning algorithm Decentralized Distributed PPO released to the public
• Added Visual Language Navigation Task & Dataset with shortest path example and tests
• Added Redwood noisy depth and RGB noise models to support sim2real direction.
• First part of Sensor API redesign: extract out visual parts of Sensor into VisualSensor, refactor RenderCamera to inherit from MagnumCamera
• The shortest path can be calculated for different agent sizes with programmatic Navmesh recomputation
• Implemented support of Gibson Semantics scene meshes and information from 3DScenegraph to extend datasets support
• Integrated clang-tidy to improve C++ coding practices, prevent memory leaks and bugs with multiple checks
• Implemented FPS counter using imgui to make performance improvement work easier
• Saved 25% of memory usage on vertex position buffer (PTex Mesh)
• Added preliminary support of continuous documentation build, added Resource manager docs
• Improved JS testing framework: setup Visual Regression Testing
• Added test coverage measurement and uploading to codecov for C++, Python, JavaScript
• TypeScript support for JavaScript build
• Added initial support for object navigation task in JavaScript
• Allow free form arbitrary yaml configs
• Numerous bug fixes, documentation and performance improvements.