A. 方向画像与考察重点
智能无人系统(人工智能)是杭州电子科技大学卓越学院二次选拔的核心方向之一,强调"会感知、能决策、可自主行动"的软硬件一体能力。该方向面向无人车、无人机、智能机器人等载体,融合人工智能、自动控制、计算机视觉与嵌入式系统,培养既懂算法又懂系统的复合型拔尖人才。面试并非单纯考查背诵,而是通过概念辨析、系统思维与项目经历,判断考生是否具备跨学科的可塑性与工程落地的潜质。
从培养模式看,卓越学院实行"一制三化"——导师制、小班化、个性化、国际化,为学生配备学术导师、提供科研早进实验室的机会,并以双语/全英文课程与海外交流拓展国际视野。智能无人系统实验班通常依托学校自动化(控制科学与工程)、人工智能相关学院的师资与平台,强调"感知—定位—规划—控制"的完整技术链。因此面试常把基础问题与具体载体(如无人机避障、无人车循迹)结合,例如"SLAM 为什么是无人系统自主性的关键""强化学习如何用于路径决策"。
本方向常见考察角度包括:人工智能与无人系统的关系、机器学习/深度学习基础、计算机视觉(检测、分割、识别)、传感器与多源融合、定位与 SLAM、路径规划与运动控制(PID、A*)、强化学习决策,以及 AI 安全与伦理、个人科创项目。建议考生既能讲清原理,也能用"我曾做过的某个小项目"把知识串起来,并主动呼应卓越学院"导师制+项目制"的培养特色。
关键词 智能无人系统 · 面试常考维度
- 方向定位:无人车/无人机/机器人 + AI 赋能的自主系统
- AI 基础:监督/无监督/强化学习、过拟合、CNN、Transformer
- 感知:计算机视觉、目标检测、激光雷达与摄像头
- 定位:SLAM、多传感器融合、状态估计
- 规划与控制:路径规划 A*/RRT、PID 与反馈控制
- 决策:强化学习 MDP、行为决策
- 系统观:感知—定位—规划—控制闭环
- 素养:AI 安全伦理、项目经历、一制三化理解
B. 核心概念速查
速查表 8 条关键概念(一句话解释)
- 人工智能(AI):让机器模拟感知、推理、决策等智能行为的总称,是无人系统的"大脑"。
- 智能无人系统:在无人工干预下能自主感知环境、规划决策并执行的系统,如无人机、无人车、机器人。
- 计算机视觉(CV):用算法从图像/视频中提取信息,是无人系统的"眼睛",含检测、分割、识别。
- SLAM:同步定位与建图,无人系统在未知环境中"边走边画地图并知道自己在哪里"的核心技术。
- 路径规划:在地图中寻找从起点到目标的无碰撞最优/可行路径,如 A*、RRT 算法。
- PID 控制:按误差的比例、积分、微分调节执行器,是运动控制最基础的方法。
- 强化学习(RL):智能体通过与环境交互、最大化累积奖励来学习策略,适合决策与控制。
- 传感器融合:把激光雷达、摄像头、IMU、GPS 等多源数据结合,获得比单传感器更可靠的状态估计。
C. 面试真题(Q1–Q20)
Q1什么是人工智能?它与智能无人系统是什么关系?专业重点
EN: What is artificial intelligence? How is it related to intelligent unmanned systems?
参考回答(中文)
人工智能是研究如何让机器具备感知、推理、学习、决策等智能行为的技术总称,涵盖机器学习、计算机视觉、自然语言处理、规划控制等分支。智能无人系统则是 AI 的"物理载体"与"落地场景"——它把 AI 算法装进无人机、无人车或机器人,让其在真实物理环境中自主完成任务。简单说,AI 提供"聪明的算法",无人系统提供"能动的身体",二者是"大脑"与"身体"的关系。杭州电子科技大学卓越学院设此方向,正是看中 AI 与硬件系统融合的交叉价值,依托"一制三化"的导师制与小班化培养,把学生引向感知—决策—控制的完整能力链。
Reference Answer (EN)
AI is the broad field of making machines perceive, reason, learn and decide. Intelligent unmanned systems are AI's physical embodiment—drones, self-driving cars, robots that act autonomously in the real world. AI is the "brain" (smart algorithms); the unmanned platform is the "body". HDU's Excellence School offers this direction to cultivate the full perception–decision–control chain under its tutorial and small-class system.
备考提示:"大脑—身体"的比喻很形象,开场题答清"AI 是方法、无人系统是载体"即可,并自然带出对学校方向的了解。
Q2请说明监督学习、无监督学习与强化学习的核心区别,并各举一例。
EN: What are the core differences among supervised, unsupervised, and reinforcement learning? Give one example for each.
参考回答(中文)
三者区别在于"是否有标签"以及"学习信号来源"。监督学习用带标注的数据 (x, y) 学映射,如用标注图像训练行人检测器。无监督学习面对无标注数据,目标是发现内在结构,如用聚类把传感器数据分群。强化学习没有现成标签,智能体通过与环境交互、根据奖励信号试错优化策略,如训练无人机自主避障。简言之:监督学习"有标准答案",无监督学习"自己找规律",强化学习"靠奖励反馈"。在无人系统中,感知常用监督学习,决策控制常借鉴强化学习。
Reference Answer (EN)
Supervised learning learns a mapping from labeled data, e.g. training a pedestrian detector. Unsupervised learning finds structure in unlabeled data, e.g. clustering sensor readings. Reinforcement learning learns a policy via reward signals from environment interaction, e.g. training a drone to avoid obstacles. In short: supervised = labeled answers; unsupervised = discover patterns; RL = learn from rewards.
备考提示:本题是 AI 开场必考题,答清"有无标签、信号来源、目标"三点并各举一例,最好用无人系统举例更切题。
Q3什么是过拟合?在无人系统的模型训练中如何缓解?
EN: What is overfitting and how do we mitigate it in unmanned-system model training?
参考回答(中文)
过拟合指模型在训练集上误差很小、但遇到新数据时泛化误差大,本质是学到了训练样本中的噪声与巧合。在无人系统中,若目标检测模型只在某光照、某场景下训练好,换到夜间或雨雾天就可能失效,这就是过拟合带来的鲁棒性问题。缓解手段包括:增大并多样化数据(采集不同时段、天气的数据,做数据增强)、正则化(L2、Dropout、早停)、简化模型、以及用验证集评估真实性能。对安全敏感的无人系统,还常用仿真+实车的数据闭环持续提升泛化能力,避免"只在实验室好看"。
Reference Answer (EN)
Overfitting: low training error but poor generalization to new data. In unmanned systems a detector trained only in daytime may fail at night—a robustness problem from overfitting. Mitigate via more diverse data, augmentation, regularization (L2, Dropout, early stopping), simpler models, and validation on real-world splits. Simulation-to-real data loops also help safety-critical systems.
备考提示:必背题。能结合"无人系统跨场景鲁棒性"来讲过拟合,比泛泛而谈更显专业。
Q4智能无人系统通常由哪些核心模块组成?请用"感知—定位—规划—控制"说明。专业重点
EN: What are the core modules of an intelligent unmanned system? Explain via "perception–localization–planning–control".
参考回答(中文)
典型无人系统是一个闭环:感知(Perception)用摄像头、激光雷达等传感器获取环境信息,识别障碍物、车道线、行人;定位(Localization)结合 SLAM、GPS/IMU 确定自身位姿;规划(Planning)在高精地图或实时地图上搜索无碰撞路径并做行为决策;控制(Control)通过 PID 等算法驱动电机/舵机跟踪路径。四者形成"看—知—想—动"的循环,并不断以反馈修正。杭州电子科技大学此方向强调这条完整技术链,正是因为它决定了系统能否真正"自主"而非遥控。面试中能画出这条闭环,是系统思维的基本功。
Reference Answer (EN)
A typical system is a closed loop: Perception senses the environment (obstacles, lanes, pedestrians); Localization estimates self-pose via SLAM/GPS/IMU; Planning searches a collision-free path and makes behavioral decisions; Control tracks the path via PID etc. This "see–know–think–act" loop with feedback is what makes the system truly autonomous rather than remotely controlled.
备考提示:"感知—定位—规划—控制"是本题骨架,务必能脱口而出并简述每层职责,体现系统观。
Q5计算机视觉在智能无人系统中起什么作用?举两个具体应用。专业重点
EN: What role does computer vision play in intelligent unmanned systems? Give two applications.
参考回答(中文)
计算机视觉是无人系统的"眼睛",负责从图像/视频中提取语义信息,是感知层的主力。具体应用一:目标检测与识别,如无人车检测行人、车辆、交通标志,无人机识别地面目标或降落区;应用二:语义分割,把图像逐像素分为路面、天空、障碍物等,为路径规划和避障提供稠密环境理解。此外还有视觉 SLAM、车道线检测、深度估计等。视觉信息常与激光雷达互补——摄像头便宜、语义丰富,雷达测距精确,二者融合可靠性更高。杭电在图像处理与嵌入式视觉上有深厚积累,是该方向的重要支撑。
Reference Answer (EN)
CV is the system's "eyes", extracting semantic info from images—the backbone of perception. Application 1: object detection/recognition (pedestrians, vehicles, signs for cars; ground targets for drones). Application 2: semantic segmentation, labeling each pixel as road/sky/obstacle for planning. Vision pairs with LiDAR: camera is cheap and semantic-rich, LiDAR measures depth precisely; fusion improves reliability.
备考提示:用"眼睛"类比并给两个具体载体应用即可;能补一句"视觉+雷达融合"体现工程认知。
Q6图像分类与目标检测有什么区别?目标检测的输出是什么?
EN: What is the difference between image classification and object detection? What is the output of detection?
参考回答(中文)
图像分类只回答"整张图属于哪一类",输出一个类别标签和概率,如判断一张图里有没有猫。目标检测更进一步,要回答"图中有哪些物体、各在什么位置",输出是若干"边界框+类别+置信度"的集合(如 (x, y, w, h, class, score))。对无人系统而言,仅知道"前方有车"不够,还必须知道车在哪个位置、离多远,才能规划避让,因此目标检测比分类更关键。常见算法有单阶段的 YOLO 系列和两阶段的 Faster R-CNN,前者速度快适合车载实时,后者精度高。
Reference Answer (EN)
Classification outputs one label+probability for the whole image. Detection outputs a set of bounding boxes with class and confidence (x, y, w, h, class, score). For unmanned systems you must know where the obstacle is, not just whether it exists—so detection matters more. YOLO (one-stage, fast, real-time) vs Faster R-CNN (two-stage, accurate).
备考提示:区分点在于"有没有位置信息";能说出 YOLO/Faster R-CNN 并联系"实时性"是加分项。
Q7卷积神经网络(CNN)为什么适合处理无人系统的视觉任务?专业重点
EN: Why is CNN well-suited to vision tasks in unmanned systems?
参考回答(中文)
CNN 用卷积核在图像上滑动提取局部特征,具备两大特性:局部连接与权重共享。局部连接契合图像"邻近像素强相关"的特点,权重共享让同一特征(如边缘、角点)无论出现在图像何处都能被同一卷积核检测到,并大幅减少参数量。多层堆叠后,低层捕捉边缘纹理,高层组合出语义(车轮、行人轮廓)。这让 CNN 在目标检测、分割、视觉 SLAM 中表现优异,且可部署到车载/机载嵌入式 GPU 实现实时推理。它是现代无人系统感知网络的基石,后续也可与 Transformer 结合提升全局建模能力。
Reference Answer (EN)
CNN uses sliding kernels with local connectivity and weight sharing. Local connectivity fits images' spatial correlation; weight sharing detects the same feature anywhere and slashes parameters. Stacked layers go from edges to semantics (wheels, pedestrians). This makes CNN excellent for detection, segmentation and visual SLAM, deployable on embedded GPUs for real-time inference—the backbone of unmanned perception, often combined with Transformers.
备考提示:"局部连接+权重共享"是 CNN 本质,必说;可补"CNN 可部署到嵌入式端"呼应无人系统实时性需求。
Q8请解释 SLAM 的基本原理,为什么它对无人系统至关重要?专业重点
EN: Explain the basic principle of SLAM and why it is vital to unmanned systems.
参考回答(中文)
SLAM(同步定位与建图)指无人系统在未知环境中,一边用传感器(摄像头、激光雷达)感知周围,一边估计自身位姿、同时构建环境地图。它本质是一个"鸡生蛋蛋生鸡"的联合估计问题:定位需要地图,建图需要位姿,二者相互依赖、迭代优化。主流方法包括基于滤波(如 EKF-SLAM)和基于图优化(如激光 SLAM 的 Cartographer、视觉 SLAM 的 ORB-SLAM)。它至关重要,是因为没有 GPS 的室内、隧道、城市峡谷中,SLAM 是系统"知道自己在哪里、环境长什么样"的唯一途径,是自主导航的前提。杭州电子科技大学此方向对 SLAM 与多传感器融合尤为看重。
Reference Answer (EN)
SLAM (Simultaneous Localization And Mapping) builds a map while estimating the system's own pose using sensors like cameras/LiDAR. It is a "chicken-and-egg" joint estimation: localization needs the map, mapping needs the pose, solved iteratively. Methods: filter-based (EKF-SLAM) and graph-based (Cartographer, ORB-SLAM). It is vital because, without GPS indoors/tunnels/canyons, SLAM is the only way to know "where am I and what does the world look like"—the prerequisite for autonomous navigation.
备考提示:能说清"定位与建图相互依赖"这一核心矛盾,并举一两个算法名(ORB-SLAM/Cartographer),专业度立刻凸显。
Q9激光雷达与摄像头在环境感知上各有什么优缺点?为何要融合?
EN: What are the pros and cons of LiDAR vs camera for perception, and why fuse them?
参考回答(中文)
激光雷达直接输出精确的三维点云,测距准、不受光照影响,能可靠获取障碍物距离与形状,但成本高、点云稀疏、缺乏颜色与纹理语义。摄像头能获取丰富的颜色、纹理与语义信息(易于做分类识别),成本低、分辨率高,但受光照/天气影响大,且需通过视觉算法间接估计深度,精度有限。二者互补:摄像头认得"那是一辆车",雷达量得"它离我 12 米",融合后既知类别又知距离,显著提升检测与避障的可靠性。这正是多传感器融合在无人系统中的价值所在。
Reference Answer (EN)
LiDAR gives accurate 3D point clouds, illumination-invariant, great for distance/shape, but costly, sparse, no color/semantics. Camera gives rich color/texture/semantics, cheap, high-res, but light/weather sensitive and only indirect depth. They complement: camera says "that's a car", LiDAR says "12 m away". Fusion yields both class and distance—far more reliable detection and avoidance.
备考提示:用"雷达给距离、摄像头给语义"一句概括差异,再点出融合动机,是标准答法。
Q10什么是路径规划?A* 与 Dijkstra 算法有何区别?专业重点
EN: What is path planning? How do A* and Dijkstra differ?
参考回答(中文)
路径规划是在给定地图与起终点后,搜索一条满足约束(如无碰撞、最短/最省能量)的可行路径。Dijkstra 是一种广度优先式最短路算法,对所有方向一视同仁地扩展,保证找到全局最短路径但效率低。A* 在其基础上引入启发式函数 h(n)(如到目标的直线距离)来估计"还剩多远",优先扩展更有希望的方向,因此在多数情况下更快、且仍能保证最优。对无人车/无人机,A* 常用于全局路径搜索,再结合动态窗口法等做局部实时避障。理解"启发式让搜索更聪明"是关键。
Reference Answer (EN)
Path planning searches a feasible path (collision-free, shortest/optimal) from start to goal on a map. Dijkstra expands uniformly in all directions—optimal but slow. A* adds a heuristic h(n) (e.g. straight-line distance to goal) to prioritize promising directions, usually faster while still optimal. A* is common for global planning in cars/drones, paired with local methods like DWA for real-time avoidance.
备考提示:核心区别就在"是否用启发式";能举 A* 配 DWA 的"全局+局部"组合,体现对实际系统的了解。
Q11请解释强化学习的基本框架(MDP)及其关键概念。专业重点
EN: Explain the basic framework (MDP) of reinforcement learning and its key concepts.
参考回答(中文)
强化学习把决策问题建模为马尔可夫决策过程(MDP),由五元组 (S, A, P, R, γ) 描述:S 是状态空间,A 是动作空间,P 是状态转移概率,R 是奖励函数,γ 是折扣因子。智能体在每个状态 s 按策略 π 选择动作 a,环境转移到新状态并给出奖励 r,目标是学到最大化累积折扣奖励期望的策略。关键概念包括:价值函数 V(s)(某状态有多好)、动作价值 Q(s,a)、策略( actor)与值函数( critic)。在无人系统中,强化学习可用于无人机自主避障、无人车跟车决策等,通过仿真大量试错来学出策略,但需警惕现实部署中的安全与样本效率问题。
Reference Answer (EN)
RL models decisions as an MDP (S, A, P, R, γ): states, actions, transition probabilities, reward, discount. The agent picks action a by policy π in state s, gets reward r and next state; goal is to maximize expected cumulative discounted reward. Key concepts: value V(s), action-value Q(s,a), policy (actor) and value (critic). In unmanned systems RL enables drone avoidance or car-following, learned via simulated trial-and-error—but watch safety and sample efficiency in deployment.
备考提示:能写出 MDP 五元组并解释"策略/价值/奖励"三角,已是强化学习合格线;联系无人系统决策应用更切题。
Q12强化学习与监督学习在无人系统中有什么不同用途?
EN: How do reinforcement learning and supervised learning differ in their use for unmanned systems?
参考回答(中文)
监督学习需要"输入—正确输出"的标注数据,适合静态的模式识别任务,如用标注图像训练目标检测器、用标注轨迹训练行为克隆(模仿学习)。强化学习没有现成标签,靠与环境交互获得的奖励来学习,适合序贯决策任务,如无人车在仿真中通过"碰撞扣分、按时到达加分"逐步学会驾驶策略。简言之:监督学习解决"这是什么/怎样模仿示范",强化学习解决"下一步该怎么做才能最大化长期收益"。实际系统常混合使用——用监督学习做感知,用强化/规则做决策控制。
Reference Answer (EN)
Supervised learning needs labeled (input, output) pairs—good for static recognition like training a detector or behavior cloning from demonstrations. RL has no labels, learns from environment rewards—good for sequential decisions like a car learning to drive via "crash penalty, on-time bonus" in simulation. In short: supervised answers "what is this / imitate"; RL answers "what to do next for long-term gain". Real systems often mix both: supervised perception + RL/rule-based control.
备考提示:"感知用监督、决策用强化"的分工总结很实用,体现对系统分层架构的理解。
Q13请说明 PID 控制的基本原理,三个环节各起什么作用?专业重点
EN: Explain the basic principle of PID control and the role of each term.
参考回答(中文)
PID 是最经典的控制算法,控制器输出 u(t) = Kp·e(t) + Ki·∫e(t)dt + Kd·de(t)/dt,其中 e(t) 是设定值与反馈的误差。比例(P)环节按当前误差大小产生控制量,误差大则用力大,但仅有 P 会残留稳态误差;积分(I)环节累计历史误差,专门消除稳态误差,但过强会引振荡;微分(D)环节依据误差变化率提前抑制,起阻尼、减小超调的作用。在无人系统中,PID 广泛用于无人机姿态稳定、无人车速度/转向跟踪。调参就是权衡三者:P 管"快",I 管"准",D 管"稳"。它是每个无人系统工程师的基本功。
Reference Answer (EN)
PID: u(t)=Kp·e(t)+Ki·∫e(t)dt+Kd·de/dt, with e the error. Proportional reacts to current error (fast but leaves steady-state error); Integral accumulates past error to eliminate steady-state error (can oscillate if strong); Derivative dampens using error rate, reducing overshoot. Used widely for drone attitude and car speed/steering tracking. Tuning: P for speed, I for accuracy, D for stability—a core skill for unmanned-system engineers.
备考提示:能写出 PID 公式并说清"P 快、I 准、D 稳"三句口诀,是控制题的满分答法。
Q14反馈控制与开环控制有何区别?无人系统为什么多用闭环(反馈)控制?
EN: What is the difference between feedback and open-loop control? Why do unmanned systems prefer closed-loop control?
参考回答(中文)
开环控制不依赖输出反馈,按预定指令直接驱动执行器,结构简单但无法纠正扰动与模型误差,如定时浇花器。闭环(反馈)控制则持续测量实际输出,与设定值比较得到误差并据此修正,如恒温器。无人系统面对的是充满扰动的不确定环境——风扰、路面打滑、传感器噪声,开环会越偏越远。闭环通过"测量—比较—修正"不断纠偏,显著提升鲁棒性与精度,因此无人机定高、无人车循迹都采用反馈控制。其代价是需要可靠传感器与稳定的控制律设计。
Reference Answer (EN)
Open-loop drives actuators by preset commands without feedback—simple but can't correct disturbances (e.g. timer sprinkler). Closed-loop (feedback) measures output, compares to setpoint, and corrects error continuously (e.g. thermostat). Unmanned systems face wind, slip, sensor noise; open-loop drifts. Feedback corrects continuously, boosting robustness and accuracy—hence drone altitude hold and car lane-following use closed loop, at the cost of needing reliable sensors and stable control design.
备考提示:用"恒温器 vs 定时浇花"类比直观;强调"不确定性环境必须闭环"是工程常识。
Q15自主导航系统一般包含哪几个层次?各层如何协同?
EN: What are the typical layers of an autonomous navigation system and how do they coordinate?
参考回答(中文)
自主导航通常分三层:任务/行为决策层(决定"去哪、做什么",如超车或避让)、路径规划层(在地图上算全局路径与局部避障轨迹)、运动控制层(把轨迹转成电机/舵机指令并跟踪)。三层自上而下传递目标,自下而上回传状态。例如决策层说"前方拥堵请绕行",规划层据此重算路径,控制层精确执行转向。它与感知、定位紧密耦合:感知提供环境,定位提供自身位姿,决策规划控制据此行动,行动又产生新感知,形成持续闭环。杭州电子科技大学此方向强调这种"软硬协同"的全栈思维。
Reference Answer (EN)
Autonomous navigation has three layers: decision/behavior (where to go, what to do—overtake or yield), planning (global path + local avoidance trajectory), and control (turn trajectory into motor commands and track it). Goals flow top-down, states flow bottom-up. E.g. decision says "detour", planning reroutes, control steers. It is tightly coupled with perception and localization in a continuous loop—exactly the full-stack thinking this HDU direction emphasizes.
备考提示:结合 Q4 的"感知—定位—规划—控制",本题补上"决策层"与"层间协同",系统观更完整。
Q16无人机相比无人车,在感知与控制上有哪些独特挑战?
EN: What unique perception and control challenges do drones face compared to self-driving cars?
参考回答(中文)
无人机在空中运动,自由度更高(六自由度:三轴平移+三轴旋转),且常无规则路面约束,因此控制难度更大,对姿态稳定与抗风扰要求高;它更轻、载荷与算力有限,难以携带大体积雷达,往往更依赖视觉与轻量激光雷达。感知上,无人机视角俯视、运动快、纹理稀疏(如高空草地),容易视觉失效;且 GPS 在楼间/室内易丢失,更依赖视觉/惯性 SLAM 与光流。相比之下无人车有平整路面约束、可装更大传感器。理解"载体特性决定技术选型"是工程思维的体现。
Reference Answer (EN)
Drones have 6-DOF motion with no road constraint, so attitude stabilization and wind rejection are harder; they are light with limited payload/compute, relying more on vision and lightweight LiDAR. Perception is tough: top-down fast motion, texture-less scenes (grass from above), easy visual failure; GPS drops in cities/indoors, so visual/inertial SLAM and optical flow matter. Cars have road constraints and room for bigger sensors. "Payload shapes the tech choice" reflects engineering sense.
备考提示:能从"自由度、载荷、视角、GPS 可用性"四方面对比,说明你真正理解不同载体的工程约束。
Q17什么是多传感器融合?卡尔曼滤波在其中扮演什么角色?专业重点
EN: What is multi-sensor fusion? What role does the Kalman filter play?
参考回答(中文)
多传感器融合是把来自摄像头、激光雷达、IMU、GPS、轮速计等多源数据,在时空对齐后综合处理,得到比任一单传感器更准、更稳、更完整的环境与自身状态估计。其核心难点是各传感器噪声不同、频率不同、坐标系不同。卡尔曼滤波(KF)及扩展/无迹版本(EKF/UKF)是最经典的融合工具:它用"预测—更新"两步,把运动模型预测与传感器观测按不确定性加权融合,给出最优状态估计,广泛用于无人系统的定位与姿态解算。融合让系统在某一传感器短暂失效时仍能可靠工作,是安全冗余的关键。
Reference Answer (EN)
Multi-sensor fusion combines camera, LiDAR, IMU, GPS, wheel odometry after spatio-temporal alignment into a more accurate, robust state estimate than any single sensor. Challenge: different noise, rates, frames. The Kalman filter (and EKF/UKF) is the classic tool: a predict–update cycle fuses a motion model with observations weighted by uncertainty, giving optimal estimates—widely used for localization and attitude. Fusion provides redundancy when one sensor briefly fails.
备考提示:"预测—更新"两步 + "按不确定性加权"是 KF 精髓;能说"融合=冗余安全"呼应无人系统可靠性需求。
Q18深度学习模型在无人系统实际部署中有哪些局限?如何应对?
EN: What limitations do deep learning models have in real unmanned-system deployment, and how to address them?
参考回答(中文)
局限主要有四:一是数据依赖与泛化,模型在训练分布外(如极端天气)易失效;二是对抗脆弱,微小扰动可能误导检测;三是可解释性差,出错难溯源,安全敏感场景难采信;四是算力与延迟,大模型难直接上车载嵌入式端。应对之道:用数据增强与仿真生成多样数据提升鲁棒性;采用模型压缩(量化、剪枝、蒸馏)适配边缘算力;以多传感器冗余与规则安全层兜底;并用不确定性估计提示"何时不可信"。杭州电子科技大学重视嵌入式 AI,正是在真实约束下做算法落地的训练。
Reference Answer (EN)
Limitations: (1) data dependency—fails out-of-distribution (e.g. extreme weather); (2) adversarial fragility; (3) poor interpretability—hard to trust in safety-critical cases; (4) compute/latency on embedded edge. Responses: augmentation + simulation for robustness; model compression (quantization, pruning, distillation) for edge; multi-sensor redundancy and rule-based safety layers as backup; uncertainty estimation to signal "when not to trust". HDU's emphasis on embedded AI trains exactly this deployment mindset.
备考提示:面试官喜欢"不止会训练、还懂落地"的考生;"压缩+冗余+安全层"三点回答显得工程成熟。
Q19人工智能用于无人系统时,需要注意哪些安全与伦理问题?
EN: What safety and ethics issues should we consider when applying AI to unmanned systems?
参考回答(中文)
安全方面,无人系统一旦自主行动就关乎人身与公共安全,必须考虑失效兜底(如感知失效时紧急制动)、对抗攻击防护、以及"电车难题"式的伦理决策(如不可避免的碰撞如何选择)。伦理方面,需关注隐私(摄像头采集人脸/环境数据)、责任归属(事故谁负责:厂商/算法/用户)、算法偏见与透明可审计。工程上常用"人在回路"(human-in-the-loop)在关键决策保留人工接管,并用功能安全(如 ISO 26262 思路)与冗余设计降低风险。作为拔尖班学生,应有"技术能力越大、安全责任越大"的意识。
Reference Answer (EN)
Safety: autonomous action affects public safety, so we need fail-safe fallback (emergency brake on perception loss), adversarial robustness, and ethical dilemmas (unavoidable-collision choices). Ethics: privacy (camera data), liability (who is responsible—maker/algorithm/user), bias, transparency. Engineering uses human-in-the-loop for critical decisions, functional safety (e.g. ISO 26262 thinking) and redundancy. As an excellence-program student, internalize "greater capability, greater responsibility".
备考提示:此题考察综合素养。能提"人在回路、功能安全、隐私责任"说明你不只懂算法,还有工程伦理视野。
Q20如果让你设计一个"智能无人系统"小项目,你会怎么做?请描述思路。
EN: If asked to design a small intelligent unmanned-system project, how would you approach it? Describe your thinking.
参考回答(中文)
我会先明确"任务与载体":比如"基于视觉的自主避障小车"。接着定义闭环:感知用摄像头+YOLO 检测障碍物;定位用里程计/简单视觉里程;规划用 A* 或动态窗口法生成避障路径;控制用 PID 跟踪。然后搭建硬件(小车+树莓派/嵌入式 GPU+电机)与软件栈(ROS 常用),先在仿真(如 Gazebo)验证算法,再实车调参。过程中用"导师制"资源找老师把关,并量化指标(避障成功率、平均延迟)评估。关键是把"感知—定位—规划—控制"串成完整系统,而非只做一个孤立模型——这也正契合卓越学院项目制、小班化的培养方式。
Reference Answer (EN)
First fix the task and platform, e.g. "vision-based autonomous obstacle-avoidance car". Define the loop: perception via camera+YOLO; localization via odometry/visual odometry; planning via A*/DWA; control via PID. Build hardware (car + Raspberry Pi/embedded GPU + motors) and a ROS software stack; validate in simulation (Gazebo) then tune on the real car. Use tutorial-system resources for guidance and quantify metrics (avoidance rate, latency). The key is a complete perception–localization–planning–control system, not an isolated model—matching the project-based, small-class cultivation of the Excellence School.
备考提示:压轴题重点不在完美方案,而在"系统闭环思维 + 仿真到实车 + 量化评估 + 呼应学校培养",展现完整工程素养。
D. 自我介绍模板(3 版 · 中英)
二次选拔面试自我介绍通常限时 1–2 分钟。推荐结构:身份锚点(10s)→ 兴趣缘起(20s)→ 能力证据(40s)→ 与智能无人系统方向的匹配(20s)→ 展望收束(10s)。以下三版分别适配不同背景,请替换【】内容并按目标方向微调,切忌背诵痕迹过重。
版本一 · 偏人工智能算法(适合有编程/机器学习基础者)
各位老师好,我是【姓名】,来自【省份】,今年被杭州电子科技大学录取。我对智能无人系统的兴趣始于【具体契机:如用树莓派做循迹小车/参加机器人社团】。为靠近它,我系统学习了【Python、PyTorch、OpenCV】,并独立完成了一个【小项目:如基于 YOLO 的摄像头避障/用强化学习训练虚拟小车过迷宫】,在【某指标】上达到了【结果】。这个过程让我确信,真正吸引我的是"让机器在真实世界里自主感知与决策",而这正是卓越学院智能无人系统方向的核心。我了解到学院实行"一制三化"的导师制与小班化培养,希望能尽早进入实验室,在老师指导下把算法真正落到载体上。谢谢老师。
Hello professors, I'm 【Name】 from 【Province】, admitted to Hangzhou Dianzi University. My interest in intelligent unmanned systems began with 【e.g. a Raspberry Pi line-following car / a robotics club】. To get closer, I learned 【Python, PyTorch, OpenCV】 and built a small project: 【e.g. YOLO-based camera avoidance / RL-trained virtual car through a maze】, reaching 【result】 on 【metric】. What truly attracts me is "letting machines perceive and decide autonomously in the real world"—exactly the core of the Excellence School's Intelligent Unmanned Systems direction. I learned the school adopts tutorial and small-class cultivation ("one system, three modernizations"); I hope to join a lab early and, under my mentor's guidance, deploy algorithms onto real platforms. Thank you.
版本二 · 偏硬件与控制系统(适合有电子/嵌入式/竞赛经历者)
老师们好,我是【姓名】。如果用一件作品介绍自己,那是我在高中做的【项目:如四旋翼无人机机架/智能车巡线系统】。它源于【真实问题:想让车自己绕开障碍】,我用【单片机/树莓派 + 电机驱动 + 传感器】解决了【关键困难:姿态不稳/定位漂移】,最终【结果与影响:能自主循迹/在比赛中获奖】。做完它我意识到,我着迷的是"软硬协同让机器动起来",而智能无人系统方向恰好把 AI 算法与自动控制融为一体。我的短板是【诚实的小短板:如机器学习理论还不够系统】,但我的优势是有【自驱力/动手能力与调试耐心】的证据。进入实验班后,我最想跟随导师做的第一件事是【具体小目标:如参与 SLAM 或嵌入式视觉项目】。谢谢老师。
Hello professors, I'm 【Name】. If one project introduces me, it's the 【e.g. quadcopter frame / smart line-following car】 I built in high school. It started from 【a real problem: making a car avoid obstacles by itself】; with 【MCU / Raspberry Pi + motor drivers + sensors】 I solved 【key difficulty: unstable attitude / localization drift】, and finally 【result: autonomous tracking / an award】. I realized I love "software-hardware co-design that makes machines move", and this direction fuses AI algorithms with automatic control perfectly. My weakness is 【an honest gap: e.g. ML theory not yet systematic】, but my strength is evidence of 【self-drive / hands-on skill and debugging patience】. After joining, the first thing I'd like to do with a mentor is 【e.g. a SLAM or embedded-vision project】. Thank you.
版本三 · 偏志趣与跨学科思考(适合无突出奖项但思考有深度者)
各位老师好,我是【姓名】。把我带到这场面试的,是一个问题:【一个真诚的学科之问:如"机器怎样才算真正'看见'了世界"/"为什么让无人机自己飞比写个程序难那么多"】。为接近它,我读了【2–3 本有分量的书:如《人工智能:现代方法》选段/《机器人学导论》】,并自学了【一门课/一次开源项目复现】。这个过程让我确认,我要找的答案在智能无人系统方向——因为这里有【独特资源:导师制早进实验室/控制与 AI 交叉的师资/浙江雄厚的智能产业】。我清楚拔尖班的强度意味着什么,高中最艰难的一段【一句话抗压证据】让我相信自己能坚持。希望能获得这个把兴趣变成系统的机会,谢谢老师。
Hello professors, I'm 【Name】. What brought me here is a question: 【a sincere inquiry: e.g. "how can a machine truly 'see' the world" / "why is making a drone fly itself so much harder than writing a program"】. To approach it, I read 【2–3 substantial books: e.g. AI: A Modern Approach / Introduction to Robotics】 and self-studied 【a course / reproduced an open-source project】. This confirmed that my answer lies in the Intelligent Unmanned Systems direction—because it offers 【unique resources: early lab access via tutorial system / faculty bridging control and AI / Zhejiang's strong intelligent-industry ecosystem】. I know the excellence program is intense; the hardest period in high school 【one抗压 evidence】 convinced me I can persist. I hope for the chance to turn interest into a real system. Thank you.
E. 专业背景知识补充
杭州电子科技大学卓越学院是学校拔尖创新人才培养的特区,智能无人系统(人工智能)实验班即依托于此。学院实行"一制三化"培养模式:导师制为每位学生配备学术导师、鼓励早进实验室;小班化保证研讨与实操质量;个性化允许按兴趣定制成长路径;国际化通过双语课程与海外交流拓宽视野。这一模式旨在打破专业壁垒,把自动化、计算机、人工智能的知识打通,契合智能无人系统"软硬一体"的本质。
从学科支撑看,杭州电子科技大学在控制科学与工程、电子科学、计算机与人工智能领域积淀深厚,自动化与人工智能相关学院为实验班提供算法、控制与嵌入式平台的师资与实验室。智能无人系统方向正是这些优势的汇聚点:以人工智能为"脑"、以自动控制为"神经"、以机器人/无人机/无人车为"躯",覆盖感知、定位、规划、控制全链条。在国家把人工智能列为战略科技力量、浙江奋力建设"智能制造"与"数字经济"高地的背景下,无人机、智能驾驶、工业机器人产业在长三角异常活跃,对"既懂 AI 又懂系统"的复合人才需求旺盛。选择此方向,意味着把个人兴趣嵌入国家需求与浙江产业沃土,也意味着要在卓越学院导师制、项目制的浸润中,成长为能真正让机器"自主行动"的拔尖工程师。
备考建议:智能无人系统是典型的交叉方向,面试不只考算法,也考"系统闭环思维"与工程落地意识。建议以"感知—定位—规划—控制"为骨架梳理知识,准备 1 个能讲清"数据→模型→指标"的小项目,并主动表达对卓越学院"一制三化"与浙江智能产业的理解,会显著加分。