How robocasa/robocasa-gr1-tabletop-tasks works
Project type and purpose
• This is a Python-based simulation framework extension (“robocasa-gr1-tabletop-tasks”) built on top of the RoboCasa core and Robosuite.
• It provides modular definitions of 24 tabletop manipulation tasks for the NVIDIA GR-1 humanoid, plus tooling to download assets, demo tasks, playback demonstration trajectories, and connect to an Isaac-GR00T inference server for closed-loop evaluation.
File structure
Top-level components:
• setup.py, requirements.txt – packaging and dependencies
• README.md, LICENSE – documentation and licensing
• images/ – static assets for README
• robocasa/ – the Python package under development
Inside robocasa/:
• environments/ – task/environment definitions (tabletop subpackage)
• models/
– assets/ – static XML/YAML 3D model and scene descriptions
– fixtures/, objects/, robots/, scenes/ – code that constructs scene elements, registers them, and builds gym environments
• macros.py – common constants and macro definitions
• scripts/ – CLI entrypoints:
– download_*.py – fetch and unpack 3D assets into models/assets
– demo_task.py – spin up a Gym environment with random actions for visualization
– playback_dataset.py – replay HDF5 demonstration trajectories in simulation
– gr00t_finetune.py – script to fine-tune GR00T
• utils/ – shared utilities: camera, configuration, sampling, Gym wrappers, texture swapping
Auxiliary folders: gym_utils for Gymnasium support
README for additional insights
• Describes integration with Isaac-GR00T: run an inference server (RPC), then run a simulation client that streams observations to the model and receives actions.
• Outlines directories to be populated by download scripts (models/assets).
• Shows sequence of steps for environment creation, model inference, and evaluation.