CARLA Cooperative V2X Intersection
A virtual intersection manager that replaces traffic lights with a V2X-inspired, reservation-based protocol for autonomous vehicles, built and benchmarked in the CARLA simulator.
Architecture
Reservation-based intersection manager
- 01
Detect
Approach-radius check registers vehicles
- 02
Message
REQUEST / PERMIT / CROSSING / CLEARED mailbox
- 03
Queue
First-come-first-served arrival order
- 04
Check conflicts
Grant PERMIT only if intents do not conflict
- 05
Clear
Conflict box releases the reservation on exit
Key decisions
DECISION 01
Geometric approach radius instead of lane-level reasoning
Keeps registration O(n) and makes the manager map-agnostic.
→ Works on any CARLA map without per-map logic
DECISION 02
FCFS conflict-aware ordering instead of a space-time reservation grid
Avoids the heavier fine-grained grid used by systems like AIM.
→ O(n log n) sort per tick
DECISION 03
Coordinate vehicles instead of cycling signals
Coordination mattered more than raw speed: even at half speed it beat the signal.
→ -51% crossing time, +105% throughput
Architecture
Intersection Manager: the central decision-maker. It maintains a First-Come-First-Served queue, infers path conflicts from each vehicle's declared intent (straight/left/right), and grants PERMIT messages only when it's safe to proceed.
Vehicle Agent: built on CARLA's BasicAgent, overridden to report state continuously, obey PERMIT/stop decisions, and otherwise drive normally once cleared.
A lightweight mailbox communication layer stands in for real V2X hardware — vehicles publish REQUEST/INTENT/STATE messages, the manager responds with PERMIT/CROSSING/CLEARED — modeling the interaction pattern without needing a real DSRC/C-V2X stack.
Algorithms
Vehicle detection uses a simple geometric approach-radius check rather than lane-level reasoning, keeping registration O(n) and making the system map-agnostic.
Permission assignment is FCFS conflict-aware: vehicles are ordered by arrival time, and each is granted passage only if its declared intent doesn't conflict with currently active vehicles — an O(n log n) sort per tick, versus the much heavier fine-grained space-time reservation grid used by systems like AIM (UT Austin).
An axis-aligned conflict box around the intersection tracks each vehicle's active/cleared lifecycle — entering the box marks a vehicle active, exiting and clearing 75% of the approach radius marks it cleared and releases its reservation.
Results
Benchmarked head-to-head against a fixed-cycle traffic signal: 51% lower average crossing time (2.56s → 1.25s per vehicle), 105% higher throughput (23.4 → 48 cars/min), and 51% faster total clearance time for a 16-vehicle queue (41s → 20s).
Notably, even running vehicles at half speed, the V2X manager still cleared the queue faster than the signal-based baseline at full speed — coordination mattered more than raw vehicle speed.
Limitations
Single-intersection only, fully cooperative traffic assumed (no human drivers, pedestrians, or cyclists), and communication is modeled as instantaneous and lossless — none of the physical-layer realism of real V2X hardware.
Group project at USC with 3 teammates (Shahid Shaikh, Hitansh Surani, Harshal Dave).