A preprint reports that V3, one of four reinforcement-learning versions of an autonomous-driving system, ranked best on the reported end-of-training measures in randomized CARLA simulations. At episode 1,000, V3 had a median lane error of 0.0514, the lowest curve-oscillation rate among non-degenerate variants, at 45.89 percent, and the highest median curve exposure, a measure of progress through a bend, at 193.5 steps.
The test used synthetic driving episodes and sensor inputs generated in CARLA under varied environmental and traffic conditions. No human-participant or real-world vehicle evaluation was described, so the result is a simulator comparison rather than a road-safety finding.
How the hybrid system worked
The proposed framework combines reinforcement learning, software that learns through repeated trial and error, with rule-based planning, PID controllers that adjust control in response to error, and large language model (LLM) common-sense reasoning. At runtime, camera and LiDAR data formed a shared state for four parallel agents, and safety constraints could override rule-based, PID or PPO control.
The LLM's role was offline during development. It analyzed logged curve-steering episodes, identified failure patterns and recommended reward adjustments; human validation came before retraining. The reasoning layer also used approximately 100 common-sense rules from Claude to support runtime safety enforcement and reasoning.
Researchers compared four reinforcement-learning versions. V0 was trained without LLM guidance on data from over 1,000 CARLA episodes, while V3 followed three rounds of LLM-based feedback and refinement. Reinforcement learning was applied in three control modes - speed, curve and hybrid - with two independent PPO-based networks handling control.
Where V3 had the edge
Over the final 200 episodes, V3 again had the lowest reported median lane error: 0.0627, compared with 0.0780 for V0, 0.0733 for V1 and 0.0770 for V2. The paper described this as approximately 15 to 20 percent lower than the other versions and reported a median steering-change magnitude of 0.3119.
That lead came with a trade-off in the curve measure. V3's exposure-weighted curve-oscillation rate was 45.27 percent with a median curve exposure of 160 steps. V0's rate was lower, at 43.18 percent, but its exposure was 113 steps; V1 and V2 recorded rates of 49.18 percent and 47.61 percent, with median exposures of 99 and 135.5 steps. The authors described V3 as the strongest combined performer, cautioning that a lower oscillation rate without comparable exposure may not indicate genuine stabilization.
End-of-training behavior was summarized descriptively rather than inferentially. The study used medians, interquartile ranges, the 10th through the 90th percentiles, episode-level observations and exposure-weighted curve-oscillation rates for the final 200 episodes; it reported no confidence intervals or significance tests.
Why the result remains limited
The evaluation did not test human driving or real-world vehicles, so it does not establish road safety. It also cannot show that LLM feedback alone accounted for the differences between V0 and V3.
The agent weights and safety priorities were design assumptions for urban driving with CARLA Town02 as the primary environment. The authors note that highway or Autobahn extensions may require adjustment.
Paper data and sources
Original title: Multi-Agent Orchestration with the Common-Sense Reasoning Capabilities of LLMs for Autonomous Driving
Authors: Mehdi Azarafza, Faezeh Pasandideh, Ali Ehteshami Bejnordi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text