Return ids with visible pixels in the instance-segmentation image.

Determine turn direction

Returns: List[Dict]: ，

[ ] ( )

Convert prompt granularity to grid cells.

Convert prompt granularity to grid cells.

Horizontal distance from a point to a node's bbox surface, with point fallback.

Positions to try when the exact pre-interaction anchor is blocked. Opening doors can make the old foot position invalid in Unity. Prefer

Return whether a graph node should block first-person walking.

CSV Completed=false： （ core/agent/graph.py fail_reason ） - DONE / FAIL / max_steps - 3 step ACTION（ ）

Return True for Unity executor failures that are safe to retry once.

Launch the VirtualHome backend and return the process.

Load configuration file

Load configuration file

Load configuration file

jsonl.gz scene_index （ train split）

prior ~/.prior ， API/token

Load task configuration from a task.json file Args: task_json_path: Path to task.json file Returns:

Load task configuration from a task.json file Args: task_json_path: Path to task.json file Returns:

Load task configuration from a task.json file Args: task_json_path: Path to task.json file Returns:

Load task ids from JSON ([..] or {\"task_ids\": [...]}) or one id per line.

Load task ids from JSON ([..] or {\"task_ids\": [...]}) or one id per line.

Return a /models URL for local OpenAI-compatible base_url, else None.

Recursively merge two configuration dictionaries Args: base: Base configuration (contains default values) ove

Recursively merge two configuration dictionaries Args: base: Base configuration (contains default values) ove

Recursively merge two configuration dictionaries Args: base: Base configuration (contains default values) ove

Normalize x_display input to AI2-THOR compatible format.

， （ spatial-planning ）

Return 'success', 'failure', or None for pending/unknown.

Return True when a point is inside a node bbox after optional shrink.

Check TCP reachability for Unity endpoint.

True if agent can MoveAhead then MoveBack (navigable, not stuck).

True if agent can MoveAhead then MoveBack (navigable, not stuck).

Flush, close, and detach a FileHandler from _log.

Flush, close, and detach a FileHandler from _log.

2D

（ ）

OpenGL 3D

Refresh active first-person pitch camera pose to reduce animation-induced drift.

Return (move_dist, rotate_angle) for the action and granularity level. gran_cfg is the navigation.granularity block from config.

Return (move_dist, rotate_angle) for the action and granularity level. gran_cfg is the navigation.granularity block from config.

ProcTHOR ： 1) config.env.dataset_dir 2) PROCTHOR_DATASET_DIR 3) <project_root>/datasets/<dataset_

： （Small/Medium/Large） Small=0.25m, Medium=0.5m, Large=1m Small spatial-planning/envs/ai2thor/wrapper.py

Map target meters (or small/medium/large) to number of [WalkForward] primitives.

Restore the character anchor after VH interaction root motion. Some VirtualHome interaction animations move the character to an object

Restore pitch by switching to the nearest registered first-person camera.

Conservatively correct interaction-induced yaw drift. Some Grab animations can rotate character yaw by one or more 30-degree steps. O

Closed-loop yaw recovery with one-step turns and direction self-correction.

evaluate （ main ， try/finally ）

API/

Select a stable camera ID that best matches first-person interaction view.

Select turning waypoint. CARLA coordinate convention (UE4 left-handed): positive yaw change => clockwise => RIGHT turn ne

Refresh graph/camera once before retrying a Unity-transient interaction.