How roboverseorg/roboverse works
RoboVerse combines MetaSim’s simulator-agnostic robotics runtime with robot, scene, and task content and learning workflows. A typical path is to select a registered task, construct its scenario and simulator environment, then use the environment for policy training or evaluation; imitation-learning workflows can also collect demonstrations and train from converted datasets. The README documents broad simulator, task, RL, IL, and VLA capabilities. Sampled source confirms the manager environment and fusion orchestration; wiring for unsampled subsystems is shown only where documented.
Main entry points: sim Simulator handlers · manager_based_rv_env.py Manager RL env · pipeline.py Fusion pipeline · scenario.py Scenario config
Components 20
Simulation runtime
- Task registry
packages/metasim/metasim/task/registry.py - Scenario config
packages/metasim/metasim/scenario/scenario.py - Task environment
packages/metasim/metasim/task/base.py - Manager RL env
roboverse_learn/managers/manager_based_rv_env.py - Simulator handlers
packages/metasim/metasim/sim - Domain randomization
packages/metasim/metasim/randomization/dr_manager.py - State queries
packages/metasim/metasim/queries
Robotics content
- Robot catalog
roboverse_pack/robots - Scenes and grounds
roboverse_pack/scenes - Task content
roboverse_pack/tasks
Learning workflows
- RL training
roboverse_learn/rl - Imitation learning
roboverse_learn/il - VLA workflows
roboverse_learn/vla - Fusion pipeline
roboverse_learn/fusion/pipeline.py
Demonstration data
- Demo collection
roboverse_learn/fusion/collect.py - Dataset conversion
roboverse_learn/il/data2zarr_dp.py - IL datasets
roboverse_learn/il/datasets
External integrations
- Simulator frameworks
- Benchmark integrations
roboverse_pack/tasks
Other components
- Engineer
Connections 23
- Engineer → Task registry: selects task
- Task registry → Scenario config: resolves task
- Scenario config → Task environment: configures
- Task environment → Simulator handlers: constructs handler
- Manager RL env → Task environment: initializes via
- Manager RL env → Simulator handlers: simulates
- Robot catalog → Scenario config: configures robots
- Scenes and grounds → Scenario config: configures scene
- Task content → Task registry: registers tasks
- Manager RL env → Domain randomization: optional randomizes
- Manager RL env → State queries: reads state
- Simulator handlers → Simulator frameworks: adapts to
- Simulator handlers → RL training: supports training
- Simulator handlers → Imitation learning: supports evaluation
- Fusion pipeline → RL training: dispatches training
- Fusion pipeline → Demo collection: dispatches collection
- Fusion pipeline → Dataset conversion: dispatches conversion
- Fusion pipeline → Imitation learning: dispatches training
- Demo collection → IL datasets: writes demonstrations
- Dataset conversion → IL datasets: converts to Zarr
- Imitation learning → IL datasets: loads datasets
- VLA workflows → IL datasets: uses robot data
- Benchmark integrations → Task content: provides task ports