Return True if a result.json already exists for this task.

Check whether Unity marks the object CLOSE to the character. ``interaction_distance_meters`` is retained as an opt-in legacy/local di

Check if the traffic light controller currently indicates red.

Return True for regular scene objects (exclude rooms/characters/structures).

Return whether a node id should count as an object visible to the agent.

pygame

jsonl.gz ProcTHOR

Move up to the requested number of cells, stopping before walls.

Move up to the requested number of cells, stopping before walls.

Validate a structured VH action dict against the prompt action space.

Normalize success condition inputs to a canonical internal representation.

Parse character_cameras() output that may be list or JSON string.

Normalize configured visibility sources to an ordered unique list.

Refresh cached scene object tokens used by prompt injection.

Refresh predicates/evaluator after success-condition configuration changes.

Resolve move magnitude from explicit value or granularity (Small/Medium/Large). Small=0.25m, Medium=0.5m, Large=1m. Default when omitted: Smal

Resolve an action target id using explicit binding, replay binding, then scene lookup.

Restore character rotation using TurnLeft/TurnRight. Args: target_quat: Target quaternion [x, y, z, w] (Unity format).

Run subprocess and decode stdout/stderr as UTF-8. On Windows, ``text=True`` without reliable UTF-8 can use GBK and crash (UnicodeDecodeError)

Capture a frame from VirtualHome and save it as a PNG. Uses the first available first-person camera, or any camera if none found. Fal

/replay {2.5, 5.0, 10.0}

/replay {30.0, 90.0}

/replay {2.5, 5.0, 10.0}

Teleport to adjacent driving lane in the same direction.

Teleport turn at junction. Blocked when traffic light is red.

Import VirtualHome package; fallback to common local clone paths.

Try restoring exact rotation via move_character; verify against live graph yaw.

Add frame to video Args: frame_data: Frame data

Convert VirtualHome graph into an AI2THOR-compatible object list. This allows existing evaluators in evaluators/base.py (which expects metada

Check if there is a pedestrian ahead Returns: (has_pedestrian, distance) tuple

Check if task completed successfully (JSON present, then evaluate). Returns: True if task completed and evaluate passes.

Check if task completed successfully (JSON present, then evaluate). AI2-THOR benchmark ： JSON ， evaluate_task

Check if task completed successfully (JSON present, then evaluate). Returns: True if task completed and evaluate passes.

Check if traffic light is red Returns: True if need to stop (red light)

Check if there is a vehicle ahead Returns: (has_vehicle, distance) tuple

Clear all memory

Close environment and release resources

Close environment

Configure task parameters (generalized support for multiple tasks) Args: target_object_types: Target object type list (e.g., ["Fr

responses chat Args: conversation_history: responses Returns: chat

Archive completed task outputs without deleting existing data.

Archive completed task outputs without deleting existing data.

outputs_true Args: task_id: ID（ retry ， "ai2thor03023"） task_output_dir: （ retry ， "ai2thor0

Copy directory without deleting existing targets.

Copy directory without deleting existing targets.

Count true/false/null in CSV Completed column. Returns dict with true, false, null, total.

Count true/false/null in CSV Completed column. Returns dict with true, false, null, total.

Count true/false/null in CSV Completed column. Returns dict with true, false, null, total.

OpenAI API - API ， ： prompt ， tools Args: client: OpenAI model: conve

CSV：true / false / None( ， ) fail_reason ， failure_type null

Infer n for max_steps = 10 + 2n from common task fields.

，

Evaluate all conditions with specified logic Args: env: Environment instance metadata: Environment metadata

Execute one navigation action via teleport-style pose update. Movement actions support small/medium/large distances. Rotation actions sup

Args: action: Returns:

JSON

golden_actions

Extract model.vlm from a single-agent config file.

model.vlm

Increase system step count when all expected active agents have acted.

Parses an action string like "MoveForward(distance=10)" or "TurnLeft"

Get current speed (km/h)

2D ； inf