Source-linked AI summary
Open-Source Autonomous Driving System Analysis and Multi-Disciplinary Hardware-in-the-Loop Research Paradigm with Reinforcement-Learning Testing and Large Language Models
Dianjing Cheng, Yike Li, Lan Yang, Shan Fang, Wenjia Niu, Xiangyu Shi, Xinyi Zhao, Yunzhe Tian, XingYu Wu, Xiaoshu Cui, Yuanwan Chen, Jialu Sun, Zhongli Wang, Biao Liu, Jiaqi Yang, Jinghui Feng, Feifei Su, Juan Du, Shuangde Fang, Yi Qian, Huiyun Li, Yuansheng Liu, Peng Sun, Mingming Wan, Nan Chen, Ruipeng Gao
TL;DR
The paper addresses the difficulty of maintaining continuous, reviewable records when real-vehicle experiments, code changes, and hardware tests are distributed across teams and stages. It proposes an Apollo-on-Hongqi EV HIL framework that connects these materials with their operating conditions, and preliminary evidence shows that the linked records support cross-task review.
Problem
Shared real-vehicle research often separates vehicle runs, code modifications, hardware tests, and feedback from their operating conditions, limiting continuous and reviewable interpretation.
Method
The paper develops an Apollo-on-Hongqi EV hardware-in-the-loop framework connecting multi-vehicle experiments, repository-based code reuse, software–hardware testing, LLM evidence structuring, and RL scenario generation.
Results
Preliminary evidence organizes vehicle operation, code reuse, hardware execution, test preparation, and sensor records within one software–hardware collaborative review process.
Takeaways & Limitations
Keeping experiments connected with their original operating conditions provides a reviewable basis for Apollo-on-Hongqi EV research and later cross-task comparison.
Takeaways & Limitations
The current evidence remains preliminary and limited in scale, requiring additional vehicle records, controlled experiments, and comparative review for broader collaborative research.
Abstract
from arXiv · showhide
Open-source autonomous driving systems provide an inspectable software foundation for intelligent vehicle research. Under real-vehicle deployment conditions, the recording and review of experimental conditions are important for interpreting system behavior and reusing experimental results. However, in a shared real-vehicle environment involving multiple vehicles, task processes, code modifications, and hardware testing feedback are often distributed across different teams and experimental stages, making it challenging to maintain continuous and reviewable experimental records. To address this limitation, this paper examines an Apollo-on-Hongqi EV environment and proposes a real-vehicle experimental framework. The framework connects multi-vehicle experiments, repository-based code reuse and software-hardware testing feedback within a unified review process. Large language models and RL-based testing serve as auxiliary components for record organization, anomaly summarization, and simulation-based candidate scenario generation. Based on this setting, this paper analyzes preliminary evidence from multi-vehicle collaborative experimentation, code and experimental-skill sharing, and software-hardware collaborative testing. The analysis shows that experimental records can be examined together with their operating conditions, providing a reviewable basis for Apollo-on-Hongqi EV research.
1 Introduction
The paper identifies a gap in continuous, reviewable real-vehicle research records and proposes an Apollo-on-Hongqi EV framework linking vehicle experiments, code reuse, and software–hardware testing. Preliminary evidence is organized across collaborative vehicle operation and testing processes.
- Open-source platforms support inspectable system-level research, but real-vehicle behavior depends on coupled sensing, computation, vehicle response, and communication conditions.
- Existing real-vehicle studies often lack continuity across teams, unified sharing of code and experience, and sustained links between software analysis, hardware testing, and external feedback.
- The Apollo-on-Hongqi EV environment retains experimental materials with their production conditions, enabling later review across university–industry research activities.
- The proposed HIL framework connects multi-vehicle experimentation, Apollo code reuse, and software–hardware testing within one review process.
- LLM-assisted evidence structuring and reinforcement-learning scenario generation support traceable review, anomaly-evidence organization, and candidate-scenario preparation.
2 Related Work
Related work establishes the value of open-source platforms, simulation, hardware-in-the-loop validation, and intelligent scenario generation, while highlighting the need to retain validation evidence with its execution conditions. This paper focuses on organizing such evidence for multidisciplinary, multi-vehicle research.
- Apollo and Autoware provide open-source environments for modular development and validation, but hardware access, complex scenarios, and co-simulation constrain broader system validation.
- Comparative work identifies Apollo’s Cyber RT and shared-memory mechanisms as enabling lower latency in large-scale data transmission.
- Real-vehicle and simulation interfaces reduce validation cost, while differences in vehicle dynamics and execution delay limit direct transfer from simulation to real vehicles.
- Hardware-in-the-loop and vehicle-in-the-loop studies improve test-environment credibility by comparing virtual, hardware, and real-road execution conditions.
- LLM and reinforcement-learning methods support executable scenario construction, safety-case organization, and boundary-scenario search before real-vehicle operation.
- For continuous real-vehicle research, validation results must remain associated with vehicle conditions and software states to support later code review, testing, and cross-team examination.
3 Apollo-on-Hongqi EV Platform and Operational Evidence Chain
The Apollo-on-Hongqi EV evidence chain links vehicle operating conditions, Apollo runtime records, and post-run review materials. This linkage supports replay and interpretation of observed responses under the original execution context.
- The Hongqi EV supplies the operating context while Apollo records the runtime process associated with each observed vehicle response.
- Each run retains vehicle-side conditions together with Apollo recording and later testing materials, keeping operation, recording, and review connected.
- Recorded real-vehicle inputs and runtime responses preserve the execution context needed for later analysis of experimental results.
- Associating each run with its Apollo record enables results to be traced back to the test environment, configuration, and communication state that produced them.
- A run-level index links each retained record to its test period, software state, and scenario segment for contextual review.
- Apollo replay allows researchers to examine recorded communication after testing and check observed responses against retained runtime material.
- The evidence chain provides a concrete reference for post-run review and preparation of subsequent hardware-in-the-loop experiments.
4 Proposed Multi-Disciplinary Hardware-in-the-Loop Research Paradigm
The paradigm organizes each real-vehicle run as a traceable experimental unit linking vehicle conditions, Apollo software state, configuration, scenario, runtime records, and review information. It connects multi-vehicle collaboration, condition-dependent code reuse, and evidence-constrained software–hardware testing within a shared review process.
- Multi-Vehicle and Multi-Team Collaborative Experiment Mechanism: Each retained vehicle run is represented as a traceable unit linking vehicle identity, Apollo software state, configuration, scenario segment, runtime record, time index, and review information.These units align vehicle and team conditions while preserving the context needed for later code reuse and testing review.
- Multi-Vehicle and Multi-Team Collaborative Experiment Mechanism: Multidisciplinary and cross-vehicle observations are attached to the corresponding run, enabling replay, configuration rechecking, and abnormal-phenomenon interpretation under recorded operating conditions.Comparable runs across Hongqi EVs support explicit comparison when scenario and configuration conditions are similar.
- Multi-Vehicle and Multi-Team Collaborative Experiment Mechanism: The mechanism treats real-vehicle experiments as evidence-linked units connecting observed vehicle responses with Apollo runtime records, providing a basis for subsequent code reuse and software–hardware testing.The shared experimental problem and run conditions define the context for later review.
- Code and Experimental Skill Sharing Mechanism: LLM-assisted retrieval identifies earlier Apollo code outcomes for new vehicle tasks, while controlled vehicle-side debugging rechecks runtime paths, data links, logs, and verification records before migration.The retrieval task uses target vehicle conditions, reuse purpose, expected runtime behavior, and evidence-linked asset packages.
- Code and Experimental Skill Sharing Mechanism: Code sharing preserves software modifications together with activation conditions, execution records, validation evidence, and reuse boundaries for cross-vehicle and cross-task review.Repository materials and internal knowledge materials are connected through experimental asset packages.
- Software-Hardware Collaborative Testing Mechanism: LLM-assisted evidence structuring partitions task-output fields into traceable and unresolved content, while reinforcement learning generates constrained simulation candidates linked to original vehicle records.The resulting assessment can locate review focus in software execution, hardware operating conditions, or scenario constraints and guide subsequent simulation or HIL testing.
5 Preliminary Result Analysis
Preliminary evidence links multi-vehicle operation, code reuse, hardware-side execution, and record-constrained simulation within a shared review process. The records preserve operating conditions and support review of vehicle behavior, code activation, board inference, and DQN-generated scenarios, while broader evaluation remains preliminary.
- Multi-Vehicle Collaborative Experiment Evidence: The multi-vehicle Apollo-on-Hongqi EV setting retains vehicle conditions with control-process observations for review across comparable runs.The automated-control model provides a reference for relating control commands, vehicle response, chassis response, and abnormal phenomena.
- Code and Experimental-Skill Sharing Evidence: The code-submission case records an Apollo CenterPoint modification with its activation condition, verification procedure, and reuse boundary.The optional proposal-refinement extension remains disabled by default and preserves original Cyber RT channels and perception-message semantics.
- Edge-Side Execution Evidence: 254 of 265 sample blocks, or 95.85% of tested blocks, contained the recognized C-V2X PC5 sidelink communication signal.The board-side result documents inference completion and structured output under the recorded 1 MB SRAM and weight-file loading setting, not full deployment effects.
- Preliminary LLM-Assisted Review and DQN-Guided Scenario Generation: The retained LiDAR review record contains 600 continuous frames over 60.035 s, with 88.57% of valid points within 20 m and 89.65% between -2 m and 1 m.The review preserved the Apollo topic source, frame continuity, and spatial distribution, while excluding detection-performance claims without detection outputs.
- Preliminary LLM-Assisted Review and DQN-Guided Scenario Generation: Across five seeds, DQN produced an independently reviewable state in 99.83% of evaluation episodes versus 5.17% for random sampling.The result is preliminary evidence for the reduced record-constrained simulation pilot; separate experiments are required for the three task-specific MDP formulations.
6 Discussion and Conclusion
The paper organizes Apollo-on-Hongqi EV experiments into a traceable review process linking vehicle operation, code reuse, hardware execution, testing preparation, and sensor records. Preliminary evidence supports reviewing experimental outputs alongside their operating conditions, while the current evidence remains limited in scale.
- The framework links vehicle operation with Apollo runtime records and experimental records so real-vehicle tests can be traced and compared across tasks.
- Experimental records remain connected to their original operating conditions, allowing outputs from one stage to be checked against records from another during later review.
- Three evidence groups cover multi-vehicle use, Apollo code modification and reuse, and hardware execution with controllable perturbation inputs.
- The combined process organizes vehicle operation, code reuse, hardware execution, test preparation, and real-sensor records within one software–hardware collaborative review process.
- The current evidence is limited in scale, and broader task-complete evaluation, repeated vehicle runs, controlled perturbation trials, stronger search baselines, and wider HIL comparisons remain for future work.