Place Cameras Aimed At a Target

Purpose

Ring a single object with N cameras that each have a clear, unoccluded view of it, using the CircularCameraPlacement API of isaacsim.sensors.rtx.placement.

Prerequisites

• Running Isaac Sim 6 / Kit 110 with python_server (port 8226); see isaac-sim-remote.
• Extension isaacsim.sensors.rtx.placement enabled (action_and_event_data_generation app, or --enable).
• Stage contains the target object and a baked navmesh (or pass scope with on_navmesh=false).
• Shell env: $ISAAC_SIM_DIR, $ISAAC_LAB_DIR, $WORKSPACE_DIR (isaac-sim-orchestrator).

Limitations

• Does not build isaacsim.sensors.rtx.placement — consume the pin from source/apps/isaacsim.exp.action_and_event_data_generation.base.kit.
• Live Kit only (raycast + navmesh); no offline substitute.
• Occlusion uses target bounding boxes, not per-triangle visibility.
• Sampling is stochastic; random_seed / density / geometry change results.
• For floor-area coverage ratios (not single-object aim), use place-camera-max-coverage.

Troubleshooting

When to use this skill

How the algorithm works

1. Hide the target prim, then raycast a sweep of candidate observation directions around it, constrained by height_range, distance_range, look_down_angle_range, and yaw_ranges.
2. Restore the target's visibility, then walk each direction outward in camera_distance_step_size increments, keeping positions that are on the navmesh (and inside scope, when given).
3. Reject any pose whose occlusion ratio for the target is at or above occlusion_threshold.
4. From the surviving directions, pick the num_cameras most widely separated (math_util.select_sparse_vectors) so the views are not redundant, and author a Camera prim per pick.

Full parameter semantics and the non-obvious validation behavior live in references/aim-at-api.md; occlusion internals and the destructive refresh flag are in references/aim-at-runtime-notes.md.

Available Scripts

Running scripts

The script is a python_server payload, not a standalone program. From agent runtimes that expose skill execution helpers, invoke it with run_script():

run_script("scripts/place_camera_aim_at.py", args=["action=preflight"])

From a built tree, send it through the isaac-sim-remote client:

python skills/isaac-sim-remote/scripts/isaacsim_send.py \
  --file skills/place-camera-aim-at/scripts/place_camera_aim_at.py \
  --arg action=place --arg target_prim=/World/Pallet --arg num_cameras=3

Workflow

1. Preflight. action=preflight (the default) reports extension, stage, target, mesh count, and navmesh state without touching the stage. It never raises, so it is safe to run first.

   run_script("scripts/place_camera_aim_at.py", args=["action=preflight", "target_prim=/World/Pallet"])
   

   Expect: AIM_AT_PREFLIGHT: ext=enabled stage=ok target=ok meshes=12 navmesh=ok. Resolve anything that is not ok before continuing — the remedy lines say how.

2. Place. Give either target_prim (preferred — enables occlusion filtering) or target_point.

   run_script(
       "scripts/place_camera_aim_at.py",
       args=["action=place", "target_prim=/World/Pallet", "num_cameras=3", "height_range=2,4"],
   )
   

   Expect AIM_AT_RESULT: placed=3 ... plus one camera[i] position=... focus=... line each. Fewer cameras than requested is a hard error, not a warning.

3. Verify. Cameras land under the parent prim path (default /World/Cameras). Confirm the count and eyeball one render before trusting the placement — the occlusion test is bbox-based.

4. Iterate. Vary random_seed for a different arrangement with identical constraints; loosen occlusion_threshold or widen distance_range when placement fails for lack of surviving directions.

Calling the API directly

When you need the poses without authoring prims (spawn_camera defaults to False):

from isaacsim.sensors.rtx.placement import CircularCameraPlacement

poses = CircularCameraPlacement.circular_camera_placement(
    target_object_path="/World/Pallet",
    number_of_camera_to_select=3,
    camera_height_range=(2.0, 4.0),
    camera_distance_range=(6.0, 18.0),
    occlusion_threshold=0.4,
    spawn_camera=False,
)
if not poses:
    raise RuntimeError("placement failed; see the carb log for the reason")
for pose in poses:
    print(pose.camera_position, pose.focus_point)

poses is a list of CameraPose dataclasses with camera_position, camera_direction, and focus_point as np.ndarray. The API returns None on every failure path rather than raising — always null-check.

Related skills

• isaac-sim-remote — the transport this skill's script rides on.
• place-camera-max-coverage — the area-coverage sibling.
• isaac-camera — per-camera intrinsics, render products, AOVs, distortion.
• isaac-sim-sensor — attaching annotators and writers once cameras exist.