Causality-Aware End-to-End Autonomous Driving via Ego-Centric Joint Scene Modeling

2026-05-13 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

CaAD, a Causality-aware end-to-end Autonomous Driving framework, addresses the causal oversight in existing end-to-end autonomous driving systems that often neglect reciprocal relations between the ego vehicle and surrounding agents. This framework captures causal inter-dependencies within a shared latent scene representation to improve trajectory predictions, especially in interaction-critical scenarios. CaAD incorporates an ego-centric joint-causal modeling module that learns dependencies between the ego vehicle and relevant agents, building on a marginal prediction branch. It also employs a causality-aware policy alignment stage using joint-mode embeddings to align the stochastic ego policy with planning-oriented closed-loop feedback from traffic and map context. The framework achieved a Driving Score of 87.53 and a Success Rate of 71.81 on the Bench2Drive benchmark, and a PDMS of 91.1 on NAVSIM, demonstrating strong closed-loop planning performance.

Key takeaway

For research scientists developing end-to-end autonomous driving systems, you should prioritize incorporating causality-aware modeling to account for the reciprocal interactions between the ego vehicle and surrounding agents. This approach, exemplified by CaAD, significantly enhances trajectory prediction consistency and reliability in complex driving scenarios, leading to improved closed-loop planning performance on benchmarks like Bench2Drive and NAVSIM.

Key insights

Causality-aware modeling improves end-to-end autonomous driving by capturing ego-agent interdependencies for consistent trajectory prediction.

Principles

• Jointly model ego-vehicle and agent causal dependencies.
• Align stochastic policies with closed-loop feedback.

Method

CaAD uses an ego-centric joint-causal modeling module to learn dependencies and a causality-aware policy alignment stage with joint-mode embeddings to synchronize ego policy with planning feedback.

In practice

• Integrate causal modeling for interaction-critical scenarios.
• Utilize joint-mode embeddings for policy alignment.

Topics

• End-to-End Autonomous Driving
• Causal Modeling
• Ego-Centric Scene Modeling
• Policy Alignment
• Bench2Drive

Code references

• AIR-THU/UniV2X

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Latent-WAM: Latent World Action Modeling for End-to-End Autonomous Driving — Latent-WAM improves autonomous driving planning via efficient, spatially-aware, and dynamics-informed latent world models.
• From Correlation to Causation in Lane Change Prediction for Automated Driving: A Causal Explanation Framework — Shifts lane-change prediction from correlation to causal reasoning for enhanced interpretability and safer automated driving.
• ScenePilot: Controllable Boundary-Driven Critical Scenario Generation for Autonomous Driving — ScenePilot generates critical, physically solvable autonomous driving scenarios that expose autonomy stack failures.
• Beyond Waypoints: A Trajectory-Centric Waypointing Paradigm for Vision-Language Navigation — Grounding waypoints in executable trajectories resolves planning-control inconsistencies and improves reachability in Vision-Language Navigation.
• Scaling Self-Play for End-to-End Driving — Self-play DAgger training in a high-throughput simulator enables scalable, robust end-to-end autonomous driving from pixels.
• A Generative Model for Closed-Loop Microsimulation of Signalized Intersections — Actor-centric generative model Enactor enables stable, realistic closed-loop microsimulation of signalized intersections by learning from its own predictions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.