Project Instinct (Instinct-Level Intelligence)

Introducing Project-Instinct, an Instinct-Level whole-body control toolkit for legged robots (especially humanoids). It integrates all features needed to train and deploy your perceptive whole-body control algorithm on a real robot, with plug-and-play modules.
To build your training environments, grouped ray-caster cameras, noise model pipelines for depth camera systems, virtual obstacle and penetration rewards for redundancy behavior and large-scale motion reference managers are at your disposal. Additional Performance Monitor manager helps you count training performance without worrying about any logic in critical training conponents.
To train your algorithm, AMP with additional loss terms, automatic encoder modules, distillation with DAgger and multi-process synchronizing support are with the bink of a config. More importantly, whenever you want to test your new algorithm, you don't need to change any of the learning code to plug your implementation into the system.
To deploy your algorithm, it shall also be simple and reusable. You only need to copy your exported training log directory to your robot, with minor customization. Everything should be ready-to-go.
More importantly, don't worry about IsaacLab/IsaacSim updates. InstinctLab is completely separate but also compatible with IsaacLab design philosophy. Everything about updates is in control.
Please visit the project website at https://project-instinct.github.io

This project design enables multi-projects running in parallel. As an example, we introduce Deep Whole-body Parkour and Hiking in the Wild, in https://project-instinct.github.io/de... and https://project-instinct.github.io/hi...