wtf is dexverse?

DexVerse is a research benchmark for teaching robots to perform dexterous, hands on manipulation tasks on a tabletop, such as picking up and handling objects with articulated robot hands. It is built as an extension on top of NVIDIA Isaac Sim and Isaac Lab, which are simulation frameworks used in robotics research. The project comes from a group of university researchers at UNC Chapel Hill, the University of Hong Kong, and UC Berkeley, and its goal is to let different robot hands and different manipulation tasks be tested against the same set of standardized environments. The repository is organized as an installable Python package under source/dexverse, which holds the actual task definitions, robot hand configurations, teleoperation device support, and shared utilities, plus a separate scripts folder with runnable entry points such as listing available tasks, running dummy test agents, recording human demonstrations through VR teleoperation, and a debug interface for manually moving robot joints. Large files like robot hand models, object meshes, and demonstration recordings are not stored in the repository itself. Instead they are downloaded separately from a gated Hugging Face dataset after the user logs in and accepts the dataset terms. Getting this running is a significant undertaking. It requires installing NVIDIA Isaac Sim and a specific version of Isaac Lab into a conda environment, along with matched versions of PyTorch built for CUDA, and several manual dependency fixes for known conflicts between these tools. Only after Isaac Lab is fully installed can the DexVerse package itself be installed as an extension. A GPU capable of running Isaac Sim is assumed throughout. As of this README, the project has shipped its initial task suite, robot and object assets, and its teleoperation and data collection tooling with documentation. Baseline environment demonstrations and baseline training code, a full demonstration dataset for the Shadow Hand, and support for additional robot hand types beyond the currently released Shadow Hand are all listed as still in progress. This project is aimed at robotics and machine learning researchers working on manipulation, not general software developers.