ToF sim-to-real — domain randomization noise sweep (research)

Context

The RL policy trains on Gazebo’s gpu_lidar-simulated TF-Luna. On the real TF-Luna A03:

• ±6 cm typical accuracy at ≤8 m.
• Amplitude-dependent noise under low contrast / overexposure.
• Periodic dropouts during servo motion (DroneBot forum).

Without noise injection in sim — the policy overfits to a “clean” sensor and produces brittle behavior on the real drone.

Final recommendation — noise injection in gpu_lidar

<sensor name="tf_luna_sweep" type="gpu_lidar">
  <lidar>
    <range>
      <min>0.2</min>
      <max>8.0</max>
      <resolution>0.01</resolution>
    </range>
    <noise>
      <type>gaussian</type>
      <mean>0.0</mean>
      <stddev>0.05</stddev>      <!-- 5 cm σ — baseline real-world TF-Luna -->
    </noise>
  </lidar>
</sensor>

Domain randomization sweep (from RL training results)

σ=0.05 m — optimum from rl-lab empirical RL experiments (see rl-lab/docs/dev-log/21-sim-to-real-noise-sweep-vl53_0-critical_HUMANED.md).

What TF-Luna behaves like in the real world

Additional randomizations for RL

1. Random rotation — sensor-frame rotation ±2° random per episode (balances “arm seats perfectly” against reality).
2. Random offset — Z calibration offset ±10 cm per episode (balances calibration drift against noise).
3. Random dropouts — p_dropout=0.05 per ray in LaserScan (mimics servo + bus issues).

Related nodes

• TF-Luna driver — health flag for discarding unreliable frames.
• LiDAR observation (1D) — how 1D LaserScan becomes a policy input vector.

Sources

1. DroneBot forum — real-world TF-Luna + servo artifacts.
2. Benewake TF-Luna A03 datasheet — accuracy specs.
3. rl-lab dev-log 21 — noise sweep VL53_0 critical (parallel research at rl-lab).
4. OpenAI Spinning Up — Sim-to-Real — general methodology.