CARGO: Causal Attribution and Responsibility-Guided Preventive Safe Multi-Agent Reinforcement Learning for Cooperative Driving

Jingjia Teng; Yang Li; Wenjie Huang

cs.AINov 6, 2026

CARGO: Causal Attribution and Responsibility-Guided Preventive Safe Multi-Agent Reinforcement Learning for Cooperative Driving

Authors: Jingjia Teng, Yang Li, Wenjie Huang

Organizations: State Key Laboratory of Advanced Design and Manufacturing Technology for Vehicle, College of Mechanical and Vehicle Engineering, Hunan University, Changsha 410082, China.

Abstract

This study presents Causal Attribution and Responsibility-Guided Optimization (CARGO), a multi-agent reinforcement learning framework for cooperative on-ramp merging in mixed traffic. A shared latent-context posterior and an action-conditioned response model estimate directed behavioral influence, with posterior uncertainty propagated into interaction-edge selection. Complementary attribution mechanisms couple signed influence with marginal team utility for cooperative reward shaping and assign responsibility from positive excess tail risk relative to a safe reference policy. Conditional diffusion generates future responses for action-dependent risk estimation. Responsibility-guided preventive optimization calibrates agent-specific risk-reduction requirements to local capacity within a policy trust region, using reserved safety margins, adaptive requirement scaling, and joint feasibility verification. Theoretical analysis establishes risk monotonicity and sufficient local conditions for non-decreasing environmental return under explicit assumptions. Experiments in highway-env across three traffic densities, together with controlled graph-recovery and attribution studies, support the effectiveness of the framework. Under high-density traffic, CARGO increases the success rate by 66.04% and reduces the crash rate by 56.23% relative to the best evaluated baseline for each metric.

Figures & tables

Fig. 1: Overview of CARGO. Interaction histories are encoded into a shared latent posterior. Multiple posterior samples support uncertainty-aware influence estimation and dynamic causal graph construction. Causal contribution combines behavioral influence with team utility to shape cooperative rewards, while causal responsibility quantifies avoidable tail risk and determines policy correction intensity and direction. An intermediate policy is accepted if its estimated costs preserve the reserved safety margins; otherwise, preventive projection enforces safety constraints and feasible responsibility-based risk-reduction requirements.

Fig. 2: Responsibility-guided preventive policy optimization. An intermediate policy is obtained by maximizing the shaped reward within a KL trust region. (a) The policy is accepted if its allowance-adjusted surrogate safety costs satisfy the reserved margins. (b) Otherwise, responsibility-based risk-reduction requirements are imposed during policy projection.

Fig. 3: Comparative study under low-, medium-, and high-density traffic. CARGO achieves higher or comparable rewards while generally maintaining lower costs than the baselines, demonstrating a favorable balance between task performance and safety.

Traffic Density	Method	Success Rate	Crash Rate	Merging Time	Average Velocity	Min TTC	TTC Violation Rate
Traffic Density	Method	(%) $\uparrow$	(%) $\downarrow$	(s) $\downarrow$	(m/s) $\uparrow$	(s) $\uparrow$	(%) $\downarrow$
Low-density Traffic	SCIC [ 11 ]	$60\pm 1.73$	$22.67\pm 4.04$	$10.83\pm 0.66$	$21.96\pm 0.22$	$1.86\pm 0.11$	$3.99\pm 0.43$
	Causality Aware [ 4 ]	$79.33\pm 3.21$	$15.67\pm 3.51$	$\underline{9.53\pm 0.06}$	$\underline{24.80\pm 0.25}$	$\underline{3.33\pm 0.24}$	$\underline{0.60\pm 0.23}$
	MAACKTR [ 33 ]	$58.67\pm 4.04$	$30.67\pm 3.21$	$11.26\pm 0.25$	$23.47\pm 0.11$	$1.89\pm 0.04$	$5.90\pm 0.58$
	MAPPO [ 36 ]	$78.67\pm 3.06$	$10.67\pm 3.06$	$10.57\pm 0.23$	$23.22\pm 0.19$	$2.99\pm 0.11$	$1.67\pm 2.05$
	Safe-MAA2C [ 7 ]	$\underline{83.33\pm 3.51}$	$\underline{2.33\pm 1.53}$	$10.08\pm 0.12$	$23.99\pm 0.10$	$2.90\pm 0.12$	$1.44\pm 0.23$

Table 1: Comparative study under different traffic densities.

Traffic Density	Algorithm	Success Rate	Crash Rate	Merging Time	Average Velocity	Min TTC	TTC Violation Rate
Traffic Density	Algorithm	(%) $\uparrow$	(%) $\downarrow$	(s) $\downarrow$	(m/s) $\uparrow$	(s) $\uparrow$	(%) $\downarrow$
Low-density Traffic	CARGO w/o CC	$68.00\pm 4.00$	$9.67\pm 2.89$	$10.58\pm 0.05$	$22.77\pm 0.20$	$2.77\pm 0.19$	$2.53\pm 0.71$
	CARGO w/o CG	$57.67\pm 4.16$	$\underline{5.33\pm 2.08}$	$12.46\pm 0.24$	$21.58\pm 0.12$	$3.06\pm 0.11$	$\underline{0.71\pm 0.32}$
	CARGO w/o CR	$59.33\pm 1.53$	$7.00\pm 1.00$	$11.96\pm 1.43$	$16.17\pm 0.17$	$2.83\pm 0.30$	$1.89\pm 0.60$
	CARGO w/o CR-PO	$\underline{86.00\pm 4.58}$	$7.67\pm 4.51$	$11.09\pm 0.20$	$23.00\pm 0.10$	$\underline{3.19\pm 0.13}$	$1.65\pm 0.30$
	MAPPO	$78.67\pm 3.06$	$10.67\pm 3.06$	$\underline{10.57\pm 0.23}$	$\underline{23.22\pm 0.19}$	$2.99\pm 0.11$	$1.67\pm 2.05$

Table 2: Ablation study under different traffic densities.

Fig. 4: Causal graph recovery under shared latent confounding, where the current actions of vehicles A and B influence the next actions of vehicles B and C, respectively. Rows denote source agents whose current actions exert influence, and columns denote target agents whose next actions are affected. Posterior-Mean Modeling learns the latent posterior but uses only its mean for graph construction, whereas No-Latent Modeling omits latent confounder representation. Entries show edge selection frequencies, ideally 1 for true edges and 0 for non-edges; gray diagonal entries are excluded. As $\rho$ decreases and direct interactions weaken, CARGO suppresses spurious edge selection more effectively than both variants while retaining most true dependencies.

Algorithm	Success Rate	Crash Rate	Merging Time	Average Velocity	Min TTC	TTC Violation Rate
Algorithm	(%) $\uparrow$	(%) $\downarrow$	(s) $\downarrow$	(m/s) $\uparrow$	(s) $\uparrow$	(%) $\downarrow$
Causal Influence	$\underline{34.67\pm 3.51}$	$\underline{12.33\pm 6.03}$	$12.04\pm 0.27$	$\underline{22.15\pm 0.26}$	$\underline{1.27\pm 0.11}$	$\underline{5.33\pm 1.05}$
Team utility	$16.67\pm 1.15$	$17.67\pm 2.08$	$\underline{11.27\pm 0.03}$	$19.73\pm 0.38$	$1.01\pm 0.02$	$9.10\pm 0.05$
CARGO	$\bm{50.67\pm 1.53}$	$\bm{2.67\pm 0.58}$	$\bm{11.24\pm 0.32}$	$\bm{22.83\pm 0.10}$	$\bm{2.47\pm 0.16}$	$\bm{1.38\pm 0.43}$

Table 3: Validation of the causal contribution component through comparison with causal influence and team utility under medium-density traffic.

Algorithm	Success Rate	Crash Rate	Merging Time	Average Velocity	Min TTC	TTC Violation Rate
Algorithm	(%) $\uparrow$	(%) $\downarrow$	(s) $\downarrow$	(m/s) $\uparrow$	(s) $\uparrow$	(%) $\downarrow$
CARGO + MAPPO-Lagrangian [ 15 ]	$30.33\pm 2.52$	$22.67\pm 3.21$	$\underline{9.81\pm 0.42}$	$\underline{24.04\pm 0.07}$	$1.32\pm 0.10$	$14.36\pm 0.98$
CARGO + PCPO [ 35 ]	$\underline{56.67\pm 4.93}$	$\underline{16.67\pm 4.16}$	$10.10\pm 0.33$	$23.80\pm 0.13$	$\underline{3.38\pm 0.16}$	$\underline{1.72\pm 0.17}$
CARGO	$\bm{88.00\pm 1.00}$	$\bm{2.00\pm 1.00}$	$\bm{9.47\pm 0.21}$	$\bm{24.85\pm 0.08}$	$\bm{3.70\pm 0.22}$	$\bm{0.48\pm 0.05}$

Table 4: Comparison of CARGO with MAPPO-Lagrangian [ 15 ] and PCPO [ 35 ] under low-density traffic.

Safe MARL

See all 9 papers

C $^2$ T: Captioning-Structure and LLM-Aligned Common-Sense Reward Learning for Traffic--Vehicle Coordination
Apr 10, 2026Yuyang Chen, Kaiyan Zhao, Yiming Wang +3Traffic Signal Control Multi-Agent Coordination
The State Key Laboratory of Internet of Things for Smart City, University of Macau · Wuhan University · Computer and Information Science, University of Macau +1
SCORP: Scene-Consistent Multi-agent Diffusion Planning with Stable Online Reinforcement Post-Training for Cooperative Driving
Apr 13, 2026Haojie Bai, Aimin Li, Ruoyu Yao +5Autonomous Driving Planning Autonomous Driving
the School of Information Science and Technology, Harbin Institute of Technology (Shenzhen), Shenzhen 518071, China · the Middle East Technology University (METU), Ankara, 06800, Turkiye · the Robotics and Autonomous Systems Thrust, Hong Kong University of Science and Technology (Guangzhou), Guangzhou 511453, China +1
Offline Multi-Agent Reinforcement Learning with a Physics-Informed World Model for Cooperative Mixed Traffic Control
Aug 18, 2026Lu Liu, Chi Xie, Xi XiongIntelligent Transportation Systems Offline MARL
College of Transportation, Tongji University Shanghai, China, 201804

Apr 10, 2026J/K move · Enter open · S save

C $^2$ T: Captioning-Structure and LLM-Aligned Common-Sense Reward Learning for Traffic--Vehicle Coordination

Yuyang Chen, Kaiyan Zhao, Yiming Wang, Ming Yang+2

The State Key Laboratory of Internet of Things for Smart City, University of Macau · Wuhan University · Computer and Information Science, University of Macau+1

Traffic Signal Control Multi-Agent Coordination Cooperative MARL LLM-Guided RL

Open paper AskPDF arXiv

State-of-the-art (SOTA) urban traffic control increasingly employs Multi-Agent Reinforcement Learning (MARL) to coordinate Traffic Light Controllers (TLCs) and Connected Autonomous Vehicles (CAVs). However, the performance of these systems is fundamentally capped by their hand-crafted, myopic rewards (e.g., intersection pressure), which fail to capture high-level, human-centric goals like safety, flow stability, and comfort. To overcome this limitation, we introduce C2T, a novel framework that learns a common-sense coordination model from traffic-vehicle dynamics. C2T distills "common-sense" knowledge from a Large Language Model (LLM) into a learned intrinsic reward function. This new reward is then used to guide the coordination policy of a cooperative multi-intersection TLC MARL system on CityFlow-based multi-intersection benchmarks. Our framework significantly outperforms strong MARL baselines in traffic efficiency, safety, and an energy-related proxy. We further highlight C2T's flexibility in principle, allowing distinct "efficiency-focused" versus "safety-focused" policies by modifying the LLM prompt.

Abstract

Safe MARL

Explore similar work

C2^22T: Captioning-Structure and LLM-Aligned Common-Sense Reward Learning for Traffic--Vehicle Coordination

SCORP: Scene-Consistent Multi-agent Diffusion Planning with Stable Online Reinforcement Post-Training for Cooperative Driving

Offline Multi-Agent Reinforcement Learning with a Physics-Informed World Model for Cooperative Mixed Traffic Control

C $^2$ T: Captioning-Structure and LLM-Aligned Common-Sense Reward Learning for Traffic--Vehicle Coordination