How stanfordvl/behavior-1k works
BEHAVIOR-1K is an embodied-AI simulation benchmark for household activities. Its central user workflow is an agent interacting with a simulated environment and being assessed on task outcomes. The tree also shows task-language semantics, evaluation and data interfaces, and a substantial asset pipeline. The README establishes the benchmark scope; most implementation excerpts are partial or unrelated, so uncertain runtime wiring is omitted.
Main entry points: env_base.py Environment API · behavior_task.py Activity task · simulator.py Simulator · activity.py BDDL semantics
Components 20
Simulation runtime
- Environment API
OmniGibson/omnigibson/envs/env_base.py - Simulator
OmniGibson/omnigibson/simulator.py
Task semantics
- Activity task
OmniGibson/omnigibson/tasks/behavior_task.py - BDDL semantics
bddl3/bddl/activity.py - Predicates
bddl3/bddl/condition_evaluation.py - State transitions
OmniGibson/omnigibson/transition_rules.py - Reward functions
OmniGibson/omnigibson/reward_functions/reward_function_base.py - Termination checks
OmniGibson/omnigibson/termination_conditions/termination_condition_base.py
Embodied world
- Scene model
OmniGibson/omnigibson/scenes/scene_base.py - World objects
OmniGibson/omnigibson/objects/usd_object.py - Object states
OmniGibson/omnigibson/object_states/object_state_base.py - Particle systems
OmniGibson/omnigibson/systems/system_base.py - Robot embodiment
OmniGibson/omnigibson/robots/robot.py - Robot controllers
OmniGibson/omnigibson/controllers/controller_view.py - Sensors
OmniGibson/omnigibson/sensors/sensor_base.py
Evaluation and data
- Benchmark evaluation
OmniGibson/omnigibson/eval/evaluator.py - Data wrappers
OmniGibson/omnigibson/envs/data_wrapper.py
Asset preparation
- Asset pipeline
asset_pipeline/b1k_pipeline - Scene conversion
asset_pipeline/b1k_pipeline/usd_conversion/usdify_scenes.py
Other components
- AI agent
Connections 18
- AI agent → Environment API: acts in
- Environment API → Simulator: steps
- Environment API → Activity task: runs activity
- Environment API → AI agent: returns observations
- Activity task → BDDL semantics: uses activity definitions
- BDDL semantics → Predicates: defines predicates
- Activity task → State transitions: applies transitions
- Activity task → Reward functions: computes reward
- Activity task → Termination checks: checks completion
- Simulator → Scene model: simulates
- Scene model → World objects: holds
- World objects → Object states: expose state
- Simulator → Particle systems: updates
- Robot embodiment → Robot controllers: dispatches control
- Simulator → Sensors: produces readings
- Benchmark evaluation → Environment API: evaluates episodes
- Data wrappers → Environment API: wraps interaction
- Asset pipeline → Scene conversion: converts assets