How physical-superintelligence-lab/simple works
SIMPLE is a simulation platform for robot policy learning and evaluation, especially humanoid loco-manipulation. A typical workflow defines a task and robot, builds a randomized scene, runs it in a simulator, and uses teleoperation or a policy to produce actions and demonstrations; evaluation runs policies against simulation and records results. The graph also shows planning, dataset handling, and the documented client-server evaluation mode. Edges reflect the excerpts, SOURCE INDEX references, or README; unsampled internals are represented only where the tree or README establishes a capability.
Main entry points: task.py Task definition · g1_inspire_wholebody.py Robot models · teleop_wbc.py Teleoperation · env_runner.py Evaluation runner
Components 21
Task and world
- Task definition
src/simple/core/task.py - Scene randomization
src/simple/dr/manager.py - World layout
src/simple/core/layout.py - Scene assets
src/simple/scenes/scene_manager.py - Robot models
src/simple/robots/g1_inspire_wholebody.py - Sensor configuration
src/simple/sensors/config.py
Simulation runtime
- Simulation environment
src/simple/envs/base_dual_env.py - Isaac Sim backend
src/simple/engines/isaacsim.py - MuJoCo backend
src/simple/engines/mujoco.py
Control and planning
- Policy and agents
src/simple/agents/base_agent.py - Teleoperation
src/simple/cli/teleop_wbc.py - Robot controllers
src/simple/robots/controllers/eef.py - Motion planning
src/simple/mp/curobo.py - Grasp estimation
src/simple/grasps/bodex.py
Data and evaluation
- Data collection
src/simple/cli/datagen.py - Dataset formats
src/simple/datasets/lerobot.py - Dataset processing
scripts/postprocess_psi0.py - Evaluation runner
src/simple/evals/env_runner.py - Evaluation metrics
src/simple/evals/metrics.py
Other components
- Operator
- Policy server
Connections 25
- Operator → Teleoperation: operates
src/simple/cli/teleop_wbc.py - Operator → Data collection: starts collection
- Task definition → Scene randomization: resets with
src/simple/core/task.py - Task definition → World layout: builds
src/simple/core/task.py - Task definition → Robot models: configures
src/simple/core/task.py - Task definition → Sensor configuration: configures
src/simple/core/task.py - Scene randomization → World layout: randomizes
src/simple/dr/manager.py - Isaac Sim backend → Task definition: runs
src/simple/engines/isaacsim.py - MuJoCo backend → Task definition: runs
src/simple/engines/mujoco.py - Teleoperation → Policy and agents: uses
src/simple/cli/teleop_wbc.py - Teleoperation → Robot controllers: commands
src/simple/cli/teleop_wbc.py - Teleoperation → Simulation environment: operates
src/simple/cli/teleop_wbc.py - Robot models → Robot controllers: uses
src/simple/robots/g1_inspire_wholebody.py - Robot models → MuJoCo backend: uses policy model
src/simple/robots/g1_inspire_wholebody.py - Motion planning → Robot models: plans for
src/simple/mp/curobo.py - Motion planning → Task definition: reads task
src/simple/mp/curobo.py - Motion planning → World layout: reads layout
src/simple/mp/curobo.py - Grasp estimation → Robot models: uses robot interface
src/simple/grasps/bodex.py - Data collection → Simulation environment: collects from
- Data collection → Dataset formats: writes episodes
- Dataset processing → Dataset formats: processes data
- Evaluation runner → Simulation environment: runs episodes
src/simple/evals/env_runner.py - Evaluation runner → Dataset formats: reads and writes
src/simple/evals/env_runner.py - Evaluation runner → Evaluation metrics: reports results
- Policy server → Evaluation runner: serves inference
README.md