Simulated RealSense D455 in MuJoCo for RTAB-Map SLAM

[zabsy]

Description

Adds src/perception/perception/slam/slam_sim_mj/, a simulated RealSense D455 that
walks a corridor loop in MuJoCo and feeds RTAB-Map, producing a 2D floor plan.

This is a simulation testbed for the real-camera procedure in slam-quickstart.md,
so the SLAM side can be developed and regression-tested without hardware: the room is
known exactly, the trajectory is repeatable, and a full run takes ~3 minutes on a
laptop with no GPU.

Design

Two halves that share no code and talk only over ROS topics:

• mj_slam_publisher.py renders the camera and publishes what a real driver
  would — /camera/camera/color/image_raw, .../depth/image_rect_raw,
  .../color/camera_info, .../imu.
• launch/slam_sim_mj.launch.py runs madgwick → static TFs → rtabmap.

Topic names, encodings and frame ids match realsense2_camera, so the SLAM half
cannot tell simulation from hardware, and the launch file runs the same
imu_filter_madgwick node with the same parameters as the real-hardware procedure.
Swapping the publisher for a real camera leaves the SLAM half untouched.

Intrinsics (fx=383.682, fy=383.165 at 640×480) are taken from a real D455 recording
rather than invented, since RTAB-Map derives scale and geometry from them.

The IMU is synthesised by differentiating the known trajectory. RTAB-Map uses an IMU
as a gravity prior in RGB-D mode; without one, odometry drifts in z on a path that is
flat by construction, and since Grid/MaxGroundHeight is 0.15 m that drift destroys
ground/obstacle classification. A physics-based IMU sensor cannot be used here — the
camera is kinematically posed, so such a sensor reads zero.

Running it

python3 -m venv --system-site-packages .venv     # repo root, once
.venv/bin/pip install mujoco imageio imageio-ffmpeg
src/perception/perception/slam/slam_sim_mj/run.sh

Outputs go to outputs/slam_mj/ (gitignored, matching outputs/rl/ and
outputs/train/). The README covers the two-terminal RViz flow and how to explore
the room interactively.

Notes for review

- The scene is generated, not committed. scene/build_room.py builds room.xml
  and its 59 textures on first run, deterministically from a fixed seed, so every
  checkout gets an identical room without carrying binaries in the repo.
- Not a ROS package. No package.xml; nothing here is built or installed by
  colcon, and the launch file is launched by path. It lives under perception/slam/
  because it is perception work, next to the quickstart it mirrors.
- 149 tests, ~0.3 s, no GPU required. The trajectory and IMU maths are pure numpy
  and tested independently — rotation and accelerometer sign errors are invisible at
  runtime and surface much later as a map that is silently tilted or mirrored.
- MuJoCo is the only new dependency, installed in a venv rather than added to any
  package manifest.