Attach collision sensor to walker (aligned with CarlaEnvWrapper / interact_carla).

scripts/carla/interact_carla.py ：sensor.other.collision

： VirtualHome ， ： "PutBack(apple, table),WalkForward" → ["PutBack(apple, table)", "WalkForward"] Ar

（ ）

Force stop and tick several frames for stabilization.

Teleport the drone to the target pose instantly (no physics delay).

Teleport the drone to the target pose instantly (no physics delay).

Map logical dual-agent IDs to AI2-THOR agent indices.

Multi-agent: map THOR agentId (0,1,...) to image subdir (agent1, agent2, ...).

Per-body folder under output_dir: agent1, agent2, ... (matches dual_agent exports).

Convert VirtualHome raw error messages to user-friendly English.

AI2-THOR spatial-planning/envs/ai2thor/wrapper.py

Translate AI2-THOR error messages to user-friendly English Args: raw_error: AI2-THOR raw error message action_name: A

Bug B fix: after reset(), check character Y. If too high, re-spawn. VHScene0 and some other scenes occasionally place the character outside t

Ensure interaction targets are visible in current camera view.

Args: max_wait_time:

，

Write a list of PIL Images as an MP4 video using cv2.

Write a list of PIL Images as an MP4 video using cv2.

AI Args: conversation_history: response_text: AI Returns:

AI Args: conversation_history: response_text: AI Returns:

Advance all traffic lights to the next state in the fixed cycle.

Apply agent-specific model config overrides from single-agent config files.

--agent1/--agent2

Keep the first system message, plus enough recent user/assistant turns and the current user message so the total number of user-image message

Keep the first system message, plus enough recent user/assistant turns and the current user message so the total number of user-image message

（ ）

（ ）

Apply initial coordinates (location and rotation).

Append deterministic traffic-light rules for all CARLA vehicle tasks.

Build complete prompt including tool descriptions and response format Args: is_first_call: Whether this is the first cal

Args: is_first_call: Returns:

Args: is_first_call: Returns:

Args: is_first_call: Returns:

Args: is_first_call: Returns:

Args: is_first_call: Returns:

Build shared context for the current agent from recent communications.

Build success condition predicate function based on success_condition configuration Returns: lambda function that receive

Build success condition predicate function based on success_condition configuration Returns: lambda function that receive

Build success condition predicate function based on success_condition configuration Returns: lambda function that receive

Build object-name guidance block for VirtualHome prompts.

Calculate angle difference between vehicle heading and target point Args: vehicle_location: Vehicle location vehicle_rotation: Ve

Capture game screenshot and encode to base64 Returns: base64 encoded image string

base64 Returns: base64

base64 Returns: base64

base64 Returns: base64

base64 Returns: base64

base64 Returns: base64

Capture front camera RGB image and return as PIL Image.

Capture front camera RGB image and return as PIL Image.

Check if current execution needs to be interrupted Returns: Interrupt reason

Check if object is in specified receptacle Args: result: Object state list expected: Expected {object_type, receptacle_type,

Check if object has specific state (isHot, isCool, isClean, etc.) Args: result: Object state list expected: Expected {object_

Use dual-agent result json directly as success source.

Use dual-agent result json directly as success source.

E:

Q:

Close the VirtualHome connection.

Compute single frame control command Returns: carla.VehicleControl

LangGraph recursion limit from per-agent max steps (10+n).