Apollo and LiDAR: Lunar Ranging & Autonomous Driving

• Apollo and LiDAR are two prominent systems that use pulsed time-of-flight measurements for both sub-millimeter lunar ranging and real-time 3D autonomous driving perception.
• APOLLO employs advanced clock synchronization and sub-10 ps calibrations to achieve millimeter-level accuracy in measuring the Earth–Moon distance.
• Baidu's Apollo integrates high-density multi-beam LiDAR with deep learning pipelines for robust object detection and precise odometry, while addressing vulnerabilities such as sensor noise and adversarial attacks.

Apollo, in the context of LiDAR research, denotes two distinct systems of international prominence: Baidu's Apollo, an industry-grade autonomous driving system leveraging real-time LiDAR-based 3D perception, and the APOLLO (Apache Point Observatory Lunar Laser-ranging Operation) experiment, the most precise lunar laser ranging facility to date. Both systems place LiDAR at the core of complex sensing, detection, and metrological tasks, yet operate at vastly different temporal, spatial, and algorithmic regimes. This article surveys both Apollo and LiDAR, detailing core principles, measurement pipelines, vulnerability analyses, and comparative metrology.

1. APOLLO: Lunar Laser Ranging as Kilometer-Scale LiDAR

The Apache Point Observatory Lunar Laser-ranging Operation (APOLLO) constitutes a kilometer-scale, time-of-flight LiDAR system engineered to measure the Earth–Moon distance by time-stamping returning photons from lunar retro-reflectors with millimeter precision. The experiment utilizes a frequency-doubled Nd:YAG oscillator (λ = 532 nm, Δt_pulse ≃ 120 ps FWHM, f_rep = 20 Hz, ~100 mJ pulse energy) transmitting through a 3.5 m telescope that functions as both beam expander and collector. Detector arrays (4×4 SPADs) are synchronously gated around the expected lunar return window (~2.5 s) to mitigate photon noise and maximize return fidelity.

Critical to APOLLO's sub-millimeter accuracy is the fusion of a 50 MHz coarse clock, a 25 ps resolution TDC, and a cesium atomic oscillator, all disciplined by an 80 MHz Absolute Calibration System (ACS) injecting sub-10 ps calibration pulses to actively monitor differential non-linearity, clock drift, and multi-photon artifacts. Range measurement entails

$R = \frac{c\,\Delta t}{2}$

with real-time corrections for atmospheric delays (Marini–Murray tropospheric models), lunar reflector tilt, Earth tides, and general relativistic Shapiro delays. APOLLO achieves median nightly precision of 1.7 mm over 15 years (cumulative photon returns O(10^2) per session), setting the standard for long-baseline laser metrology and enabling constraints on lunar ephemerides and fundamental physics (Battat et al., 2023).

2. Baidu Apollo: LiDAR-Driven Autonomous Driving Perception

Baidu Apollo is an industrial-scale Level 4 ADS platform utilizing multi-beam spinning LiDAR (e.g., Velodyne HDL-64E, 64 channels, 10 Hz, ~100–130k points per sweep) for 3D environment modeling, obstacle detection, and motion planning. The Apollo LiDAR perception stack comprises:

• Sensing: Raw time-of-flight (ToF) returns (x, y, z, intensity) from an array of rotating beams.
• Pre-processing: Transformation to global frames, ROI cropping, and partitioning into fixed vertical "pillars" or grids.
• Feature Extraction: Per-pillar statistics (max/mean height, intensity, point count, etc.) assembled into structured tensors (e.g., x ∈ ℝ^{H×W×8}).
• Machine Learning Inference: DNNs (PointPillars/PointRCNN or proprietary architectures) predict objectness, centroids, class probabilities, and fit 3D bounding boxes.
• Post-processing: Clustering, bounding box fusing, optional multi-frame tracking.

This modular processing pipeline operates on production Apollo vehicles and provides downstream inputs to trajectory prediction and motion planning modules (Pham et al., 2024, Cao et al., 2019).

3. Robustness and Vulnerability Analyses of Apollo's LiDAR Perception

3.1 Subtle Perturbation Robustness: SORBET

The SORBET framework models real-world, built-in LiDAR sensor inaccuracies (millimeter-level errors, ≤2 cm displacements, or the removal of a handful of points) and quantifies their impact on Apollo’s obstacle detection and trajectory prediction. For a given point cloud sweep $P = \{ p_1, \dots, p_N \} \subset \mathbb{R}^3$, SORBET applies:

• Jitter: $p_i' = p_i + \Delta_i$ with $\| \Delta_i \|_2 \leq \epsilon$, $$\epsilon \in \{0.01\text{m}, 0.02\text{m}, 0.05\text{m}\}$$.
• Point Removal: Global (random) or local (within ground-truth boxes).

Evaluating Apollo’s detector under local removal yields marked drops in detection metrics:

k (points/obstacle)	mAP (%)	Precision (%)	Recall (%)
0	85.4	88.2	84.1
2	78.3	83.5	77.0
5	72.1	79.0	70.4
10	60.5	68.2	59.0

Even two points removed per obstacle induce a 7.1 percentage point mAP drop; at k = 10, mAP falls by 25 points. Propagation to trajectory prediction results in a +12% to +31% increase in average/final displacement errors and a 9% higher rate of near-collisions or emergency braking in simulation (Pham et al., 2024).

3.2 Adversarial Object Attacks

Apollo's perception stack is also susceptible to shape-only adversarial objects created via gradient-based (whitebox) or evolutionary (blackbox) optimization (LiDAR-Adv). Attackers deform a mesh S = (V, F) to minimize the "positiveness" output of the detector or force a label flip, under physical constraints to ensure plausibility.

• In real-world 3D-printed tests, adversarial cubes evade Apollo’s detector in 100% of live drive-by frames, while benign cubes are detected 67% of the time.
• Success rates in simulation: 50 cm cube—62% (evolution), 71% (whitebox); angle-invariant objects can evade detection across ±10° of orientation (Cao et al., 2019).

This demonstrates that LiDAR-based detectors are not immune to geometric adversarial attacks; careful sculpting can shadow most returns or mislead feature aggregation.

4. Apollo-SouthBay Dataset: LiDAR Odometry Benchmarks

Baidu's Apollo-SouthBay dataset, collected in San Francisco Bay, is a benchmark for LiDAR odometry: sequences span urban, suburban, and highway domains with tightly coupled GPS/INS ground truth. A Velodyne HDL-64E (10 Hz) provides dense per-frame point clouds (x, y, z, intensity), which are used in both supervised and self-supervised odometry pipelines.

4.1 UnPWC-SVDLO and SelfVoxeLO Pipelines

• UnPWC-SVDLO (Tu, 2022): Employs PointPWC for multi-scale scene flow, then estimates pose via SVD-based ICP using point-to-plane residuals. Achieves 4.53% translational and 1.30°/100 m rotational RMSE on the Apollo test split—lowest among all unsupervised methods, outperforming LOAM on translation.
• SelfVoxeLO (Xu et al., 2020): Voxelizes raw sweeps to a 0.1×0.1×0.2 m sparse grid, encodes features via 3D sparse convolution, and regresses SE(3) ego-motion using 2D ResNet heads with self-supervised losses exploiting geometric consistency and per-voxel uncertainty. On Apollo-SouthBay, self-supervised SelfVoxeLO (with mapping) yields t_rel = 2.25%, r_rel = 0.25°/100 m—surpassing classic two-frame baselines and LOAM in translation.

Both approaches highlight the importance of fine geometric feature learning and explicit uncertainty modeling in large-scale odometry without reliance on pose supervision.

5. LiDAR Metrology: APOLLO Versus Terrestrial Systems

While both APOLLO and terrestrial LiDAR are based on pulsed time-of-flight measurements, key differentiators exist:

Parameter	APOLLO Lunar Ranging	Terrestrial Mobile LiDAR
Range	∼384,000 km (Earth–Moon)	meters–kilometers
Pulse energy	~100 mJ @ 20 Hz	μJ–mJ @ kHz–MHz
Timing precision	25 ps binning, <10 ps calibrations	~100 ps (∼1.5 cm range)
Target	Retroreflectors (corner cube)	Diffuse ground/building returns
Detector	SPAD array, gated	APDs or SPADs, often multi-channel
Measurement mode	Fixed pointing at known targets	Scanning/rotating, 3D point cloud build

APOLLO's architecture enables an order-of-magnitude improvement over prior LLR stations (from 10–20 mm to 1.7 mm daily uncertainty), a performance not approached by general-purpose terrestrial LiDAR, reflecting the extreme requirements of lunar geodesy and fundamental physics (Battat et al., 2023).

6. Implications, Limitations, and Prospects

Robustness gaps in Apollo’s automotive LiDAR perception—vulnerable both to subtle sensor-induced perturbations and physically realizable adversarial shapes—mandate the integration of robustness analysis such as SORBET, denoising front-ends, retraining on perturbed data, and multi-sensor fusion for safe deployment (Pham et al., 2024, Cao et al., 2019).

For LiDAR-based odometry (Apollo-SouthBay), multi-scale geometric encodings, unsupervised consistency losses, and explicit uncertainty modeling drive substantial performance gains over classic point-based and learning-based baselines. On the lunar ranging front, the APOLLO experiment establishes the upper bound for single-baseline laser time-of-flight metrology, with continued relevance for tests of gravity and planetary ephemerides.

Future work for Apollo and LiDAR includes (as suggested): robustification to sensor noise, adaptive/multi-scale voxelization, joint semantic and geometric modeling, and full-stack integration of mapping and loop closure in odometry systems, as well as continued cross-validation with additional unlabeled data (Xu et al., 2020). The precise metrological innovations of APOLLO LLR remain influential in the calibration and architecture of emerging terrestrial and planetary LiDAR applications.