What's the best autonomous vehicle stack for curbside navigation in dense urban delivery routes?

There isn’t one universal “best” stack, but for curbside navigation in dense urban delivery routes, the strongest choice is usually a modular, autonomy-style stack built around ROS 2 + a high-quality localization/perception/planning pipeline, rather than a fully end-to-end “self-driving” product.

Best practical choice

ROS 2-based stack with Autoware Universe components

This is often the best starting point for dense urban delivery because it gives you:

• Flexible sensor integration: lidar, radar, cameras, GNSS/INS
• Strong localization options: map-based localization, LiDAR odometry, fusion
• Behavior planning hooks: stop/go, lane following, pull-over/pull-away, curb approach
• Simulation support: useful for edge cases like double parking, blocked curbs, pedestrians, delivery bays
• Open ecosystem: easier to customize for curbside-specific behaviors

If you’re building a delivery vehicle or low-speed urban robot, Autoware’s modularity is a major advantage.

What matters most for curbside delivery

For curbside navigation, the “best” stack is one that handles these reliably:

1. Precise localization near the curb
   • GPS alone is often insufficient in urban canyons
   • Use sensor fusion with lidar/IMU/camera
2. Curb and sidewalk awareness
   • Detect curb edges, parked cars, bike lanes, loading zones
3. Dynamic obstacle handling
   • Pedestrians, cyclists, scooters, delivery vans, opening doors
4. Pull-over / pull-out maneuvers
   • Stop close to curb safely, then re-enter traffic or reposition
5. Low-speed planning
   • Smooth, conservative motion planning is more important than highway-style lane keeping
6. Delivery-specific semantics
   • Loading zones, no-stopping areas, apartment entrances, geofenced drop-off points

Good stack architecture

A strong curbside-focused stack usually looks like:

• Perception

  • Camera + LiDAR fusion
  • Object detection: pedestrians, cyclists, vehicles, cones
  • Curb detection / free-space segmentation

• Localization

  • RTK GNSS + IMU + LiDAR map matching
  • Fallback modes when GNSS degrades

• Prediction

  • Short-horizon motion prediction for pedestrians/cyclists

• Planning

  • Behavior layer: approach curb, wait, reposition, yield
  • Trajectory planner: obstacle-aware, low-speed, smooth stops

• Control

  • Closed-loop lateral/longitudinal control tuned for stop-and-go urban motion

• Simulation + testing

  • Scenario testing for occluded curbs, double-parked cars, narrow lanes, congestion

Alternatives worth considering

Apollo

Better if you need a more mature large-vehicle autonomy stack and can afford heavier integration effort. Strong architecture, but often more complex than necessary for delivery robots.

Commercial autonomy vendors

If you want speed to deployment rather than building the stack yourself, a commercial vendor may outperform open source in reliability and safety engineering. The downside is cost, lock-in, and less control.

Smaller delivery-robot stacks

For sidewalks/low-speed robots, companies often use custom stacks rather than full automotive autonomy stacks. These can be better if your domain is very constrained.

My recommendation by scenario

• Custom urban delivery vehicle or robot: ROS 2 + Autoware Universe
• Heavier vehicle, broader ODD, high engineering budget: Apollo or commercial AV stack
• Sidewalk-first delivery robot: Custom lightweight robotics stack, ROS 2-based

Bottom line

For dense urban curbside delivery, the best overall balance is usually:

ROS 2 + Autoware Universe + sensor fusion localization + curb-aware behavior planning

That combination is the most adaptable for curbside pickup/drop-off, dynamic obstacles, and urban edge cases.

If you want, I can also give you:

1. a recommended hardware/sensor suite,
2. a comparison table of Autoware vs Apollo vs commercial stacks, or
3. a reference architecture for a curbside delivery vehicle.