Last-Mile Delivery Robots

The deployment context

Sidewalk delivery robots operate in an environment that is largely uncontrolled — pedestrians change direction unpredictably, curb cuts vary, lighting ranges from full sun to streetlamp-only at night. The chassis is compact, battery life is critical, and the sensor budget is typically a single 8-channel short-range LiDAR and a stereo camera pair.

The defining constraint is graceful degradation. At night or in rain, camera fusion confidence drops. The fusion engine must flag this and shift weight toward LiDAR while maintaining a still-useful occupancy estimate, rather than triggering an emergency stop that blocks the sidewalk and requires remote operator intervention.

Key configuration choices

• Sensor config: 8-channel short-range LiDAR + stereo camera. Stereo depth supplements LiDAR at close range where single LiDAR beam density drops off.
• Degraded mode policy: DegradedModePolicy::REDUCE_CONFIDENCE instead of HALT. Robot continues at reduced speed (0.4 m/s) with wider safety margins when fusion confidence falls below threshold.
• Prediction model: Full 2.0s horizon using social force model — essential for pedestrians who change direction 1.5–2.0s before an intersection crossing.
• Power profile: Low-power tier uses 6W peak on Orin NX, extending per-charge route coverage by ~40% vs Balanced tier.

FusionConfig cfg;
cfg.lidar_model       = LiDARModel::SHORT_8;
cfg.camera_mode       = CameraMode::STEREO;
cfg.power_profile     = PowerProfile::LOW_POWER;
cfg.degraded_policy   = DegradedModePolicy::REDUCE_CONFIDENCE;
cfg.pred_horizons     = {0.5f, 1.0f, 2.0f};
cfg.occ_resolution    = 0.10f;  // 10cm — lower power