---
title: "CORAL: Curriculum-Optimized Reward Adaptation for LiDAR-Based Goal-Directed Urban Driving"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.14332v1"
permalink: "https://feed.astralor.com/items/itm_948a17a3d0fa0554af32ad11"
published: "2026-08-14T14:22:05.000Z"
score: 72
channel: "tech"
tags: ["CORAL","LiDAR","PPO","Reward Adaptation","Autonomous Driving","Curriculum learning"]
---

# CORAL: Curriculum-Optimized Reward Adaptation for LiDAR-Based Goal-Directed Urban Driving

## AI commentary

> The elegance of this work lies in scheduling both “what to learn” and “how to reward” simultaneously, rather than stacking network architecture. Teams working on autonomous driving or robot navigation should pay attention to its curriculum design and reward-weight scheduling, especially the generalization achieved with a low-dimensional state.

## At a glance

CORAL proposes a curriculum-optimized reward adaptation method for LiDAR-based goal-directed urban driving. It jointly advances a five-stage curriculum that gradually extends routes and tightens behavioral constraints, and a stage-aware reward whose component weights shift from mission progress toward route following, safety, smoothness, and rule compliance. The policy is a multi-stream actor-critic network trained with PPO in CARLA on a compact 99-dimensional state consisting of a polar LiDAR histogram, vehicle telemetry, ego-frame route geometry, and traffic-rule indicators. Compared with two PPO baselines under an identical protocol, CORAL reaches the goal in all twenty evaluation episodes on the longest routes under full behavioral constraints, whereas the baselines reach 5% and 10%. Factorial ablations show that neither schedule alone matches their combination; removing either lowers success and route completion, and disabling both drops success to 55%. Trained in one town, the policy transfers zero-shot to seven unseen towns, succeeding in 68-98% of episodes, with mean lateral deviation below 0.35 m.

## Key points

- Proposes CORAL, which jointly schedules a five-stage curriculum and stage-aware reward adaptation to address the learning-order problem of long-horizon driving behaviors under a fixed reward.
- Uses only a 99-dimensional state (polar LiDAR histogram + vehicle telemetry + ego-frame route geometry + traffic-rule indicators), requiring no point-cloud encoder or BEV rasterization.
- Achieves 100% success on the longest routes under full behavioral constraints, vs. 5% and 10% for two PPO baselines; ablations show neither mechanism alone matches their combination.
- After training in one town, zero-shot transfers to seven unseen towns with 68-98% success and mean lateral deviation below 0.35 m.

## Full text

Reinforcement learning is promising for autonomous urban driving, but long-horizon goal-directed navigation asks a policy to acquire several competing behaviors at once--reaching a distant goal, tracking a route, avoiding obstacles, obeying signals--and a fixed objective gives no order in which to learn them. This paper presents CORAL, which advances two schedules together: a five-stage curriculum that progressively lengthens routes and tightens behavioral constraints, and a stage-aware reward whose component weights shift emphasis from mission progress toward route following, safety, smoothness, and rule compliance as the task hardens. The policy is a multi-stream actor-critic network trained with Proximal Policy Optimization (PPO) in CARLA on a compact 99-dimensional state pairing a polar LiDAR histogram with vehicle telemetry, ego-frame route geometry, and traffic-rule indicators--no point-cloud encoder, no bird's-eye-view rasterization. Against two PPO baselines under an identical protocol, CORAL reaches the goal in all twenty evaluation episodes on the longest routes under the full set of behavioral constraints, where the baselines reach 5% and 10%; a factorial ablation shows that neither schedule alone matches their combination: removing either lowers both success and route completion, and disabling both drops success to 55%. Trained in one town, the policy transfers zero-shot to seven unseen towns, succeeding in 68-98% of episodes on routes of the same 100-150 m length, with mean lateral deviation below 0.35 m.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_948a17a3d0fa0554af32ad11