Get ProcTHOR system prompt (aligned with AI2-THOR prompt structure).

OpenAI ， API Args: response: OpenAI API api_type: API ， None Returns:

Get a top-level section of configuration Args: section: Section name, e.g., "env", "task", "model" Retur

Get a top-level section of configuration Args: section: Section name, e.g., "env", "task", "model" Retur

Get recorded video file path

Infer environment type from task id first, config as fallback.

Check if point is in polygon (Ray Casting)

Check if point is in polygon (Ray Casting)

Load init.json and normalize rotation to 30-degree discrete yaw. Returns a dict with: - init_action_strings - init_scene - init

Load task metadata from tasks/<task_id>/task.json. Returns keys: - instruction: str | None - golden_action_count: int | None -

Load task metadata from tasks/<task_id>/task.json. Returns keys: - instruction: str | None - golden_action_count: int | None -

Load metadata from mllm_base_agent/dual_agent/procthor/task_mutil_procthor/<task_id>/task.json.

Inject interact-recorded object IDs into a golden action sequence. During an interact session the controller records which specific object

（ ） decide_csv_status task_success

Normalize angle to [-180, 180].

Normalize task ID (e.g. ai2thor_04000 -> ai2thor04000, procthor_00001 -> procthor00001).

JSON Args: text: JSON ， JSON Returns: JSON ， None

Snap yaw(deg) to nearest discrete step and normalize to [-180, 180].

Convert Unity quaternion [x, y, z, w] to yaw(rad).

Read task IDs from CSV. Args: csv_path: Path to CSV. only_null: If True, only include rows with Completed null/empty. R

Record model response and decision Args: frame_number: Frame number response: OpenAI API response object

Args: frame_number: response: OpenAI API action_taken:

Args: frame_number: response: OpenAI API action_taken:

Args: frame_number: response: OpenAI API action_taken:

Args: frame_number: response: OpenAI API action_taken:

Args: frame_number: response: OpenAI API action_taken:

Reset controller state

[ ] -

Map model output to one of {2.5, 5, 10}. Accepts: - labels: small / medium / large (case-insensitive) - numeric literals: 2.5, 5, 10

Map model output to one of {30, 90}. Accepts: - labels: medium / large (case-insensitive; small is a legacy alias for medium) - numer

Save complete conversation history as multi-turn dialogue JSON format

Save complete conversation history as multi-turn dialogue JSON format

Args: task_id: ID task_output_dir: failed_logs_dir: attempt: outp

Save failed task outputs without deleting previous snapshots.

，

Save failed task outputs without deleting previous snapshots.

Retry infra/API outcomes that intentionally remain CSV null.

BFS ，

This method identifies the current stage the cube is solved to, then solves it to the next stage.

Start recording video Args: frame_data: First frame data for video parameters session_name: Session name for

Execute one step of PID calculation Args: error: Current error Returns: Control output

（ ： ``thor_agent_id`` ） Args: action_dict: ， action_type, action_name thor_agent_id:

Stop recording video

Wait for a local OpenAI-compatible API before burning a task attempt.

Convert yaw(rad) to Unity quaternion [x, y, z, w].

Initialize VirtualHome environment wrapper. Args: scene: VirtualHome scene index (0–49 correspond to different apartments).

Granular movement already represents the full step, so do not repeat it.

Create <log_dir>/agent.log and attach a FileHandler to _log.

Create <log_dir>/agent.log and attach a FileHandler to _log.