Skip to content
TheChatArchiveThe Chat Archive
o3-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

How does a LiDAR see a whole 3d outlook? Since it's spinning atop the waymo 360degrees,…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Geminigemini-2.5-flash-preview-04-17’s answer.
  1. User

    How does a LiDAR see a whole 3d outlook? Since it's spinning atop the waymo 360degrees, wouldnt it just see a 1 pixel tall representation of depth all around it?

  2. Assistant

    Think of an automotive LiDAR as a super-fast laser tape-measure that is swept across the scene in a very specific pattern.
    The trick is that it’s NOT just “one laser spinning in a flat plane.”
    Here’s how the common spinning (a.k.a. “mechanical”) units such as the old Waymo/Velodyne puck actually create a full 3-D point cloud:

    ────────────────────────────────────────────

    1. Multiple lasers, each tilted a bit ──────────────────────────────────────────── top view (cut-away) side view ┌─────────────────────────┐ vertical fan of rays │ │ / +10° (high) │ rotating head │ / +8° 🔴🔴🔴🔴🔴🔴🔴🔴 stacked emitters … │ ← 64 (or 32, 16, …) lasers │ │ | 0° └─────────────────────────┘ \ −15° 〳 −25° (low curb)

    • Every “channel” (laser + receiver pair) is fixed at its own elevation angle.
    • Typical specs:
    – 16-channel lidar ⇒ ≈ −15 ° to +15 ° vertical FOV
    – 64-channel lidar ⇒ as much as −25 ° to +10 °, 0.4 ° vertical spacing
    – Firing rate ≈ 10–20 k laser shots per channel per second.

    So, at any instant the lidar is drawing a 64-pixel-tall ring of distance measurements.

    ──────────────────────────────────────────── 2. Continuous 360° spin gives you the other axis ──────────────────────────────────────────── While those 64 (or 32, …) beams form a vertical “fan,” the whole assembly spins 5 – 20 times each second.
    As it rotates, each beam paints out a little dotted circle at its elevation angle. Put all those circles together and you get a dense cylindrical shell of points around the car.

           one full rotation → 3-D point cloud (simplified)
           ┌───────────────┐
           │   *  *  *     │  ← high-angle beam
           │ *  *   * *    │
           │*  * o  *  *   │
           │ *  *   * *    │
           │   *  *  *     │  ← low-angle beam
           └───────────────┘
    

    • At 10 Hz (100 ms per spin) a 64-channel unit can deliver ≈ 1.2 million points per frame.
    • Because everything is captured in < 100 ms, motion blur is small for normal driving speeds (and can be corrected with IMU data).

    ──────────────────────────────────────────── 3. How the 3-D coordinates appear ────────────────────────────────────────────

    1. Measure time-of-flight t of each laser pulse.
    2. Convert to distance d = (c · t)/2.
    3. Knowing the laser’s azimuth (rotation angle) and elevation angle, convert (d, az, el) ⟶ (x, y, z).
    4. Stitch all points from one rotation into a “point cloud” frame; stitch many frames over time for a map.

    ──────────────────────────────────────────── 4. What if you really only had ONE beam? ──────────────────────────────────────────── • You’d get exactly the “1-pixel-tall” ring you imagined—just a 2-D lidar.
    • Robots often mount such 2-D lidars close to the ground (e.g., indoor Roombas or warehouse AGVs).
    • To build 3-D with a single beam you must move the sensor up/down or let the vehicle’s own motion sweep the beam through space (called “push-broom” scanning). It’s slower and less dense, which is why self-driving cars prefer multi-beam designs.

    ──────────────────────────────────────────── 5. Other modern approaches (bonus) ──────────────────────────────────────────── • MEMS or “image lidar”: one laser, one detector, two tiny mirrors do a raster scan (like an old CRT).
    • Flash lidar: flood-illuminate the whole scene and use a 2-D detector array—no moving parts, but currently shorter range.
    • FMCW coherent lidar: sweeps laser frequency to get both distance and direct radial velocity (like radar Doppler). Still experimental.

    ──────────────────────────────────────────── Key takeaway ──────────────────────────────────────────── Spinning lidars use a stack of lasers (vertical resolution) + rotation (horizontal resolution) + rapid firing (dense sampling) to create a rich 3-D point cloud, not just a single ring of depths.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.