二、面试真题(Q1–Q30)
Q1请定义"具身智能",并说明它和传统软件 AI(如大语言模型)的本质区别。专业重点
EN: Define "embodied intelligence" and explain how it fundamentally differs from traditional software AI such as LLMs.
参考回答(中文)
具身智能指智能体通过具备物理形态的身体在环境中感知、交互并行动,从而习得智能的范式。它与传统软件 AI 的本质区别有三点:第一,智能来源不同——纯软件 AI 依赖文本/数据统计规律,具身智能依赖"身体经验"与物理反馈;第二,任务闭环不同——具身智能必须打通"感知→决策→控制→执行"的实时闭环,而对话模型只需生成token;第三,评价标准不同——具身任务以物理世界中的成功率为准(如抓取成功率、避障安全性),对鲁棒性、时延和安全性要求更高。简言之,具身智能是"能动起来、能在真实世界完成任务"的 AI。
Reference Answer (EN)
Embodied intelligence is a paradigm where an agent with a physical (or simulated) body perceives, interacts with, and acts in an environment to acquire intelligence. Key differences from software AI: (1) intelligence comes from physical interaction rather than text statistics; (2) it requires a real-time perceive-decide-act-control loop, not just token generation; (3) success is measured by real-world task outcomes (e.g., grasp rate, collision avoidance), demanding robustness and safety. In short, it is AI that can act and succeed in the physical world.
备考提示:考官常追问"具身"的含义,可补充:具身强调"智能离不开身体与环境的耦合",是符号主义/连接主义之外的第三条路线,答出"交互即学习"会很加分。
Q2为什么近年具身智能突然成为热点?请列举驱动因素。
EN: Why has embodied AI surged recently? List the driving factors.
参考回答(中文)
具身智能升温由四方面共同推动:一是大模型溢出效应——多模态大模型(视觉-语言模型)为机器人提供了语义理解与零样本泛化能力,催生 VLA 架构;二是算力与仿真成熟——GPU 与物理仿真(Isaac Sim、MuJoCo)让大规模训练与 Sim2Real 成为可能;三是硬件成本下降——高性价比传感器、关节模组和移动平台(无人机、轮式/四足机器人)普及;四是应用牵引——仓储物流、制造、服务机器人等真实需求迫切。同济与上海人工智能实验室在具身方向的合作,正是这一趋势下"产学研"联动的典型体现。
Reference Answer (EN)
Four drivers: (1) spillover from large models—VLMs provide semantic understanding and zero-shot generalization, leading to VLA; (2) mature compute and simulation (Isaac Sim, MuJoCo) enabling large-scale training and Sim2Real; (3) lower hardware cost—sensors, actuators, and platforms (drones, legged robots); (4) real demand from logistics, manufacturing, and service robots.
备考提示:可顺势联系同济与上海人工智能实验室的具身智能合作,体现你对学校特色的了解。
Q3专业重点专业重点请描述机器人"感知—决策—控制"闭环,并说明每一环的作用。
EN: Describe the robot perceive–decide–control loop and the role of each stage.
参考回答(中文)
自主机器人运行在"感知—决策—控制"闭环中:①感知层用相机、激光雷达、IMU 等传感器获取环境信息,并通过 SLAM/目标检测输出自身位姿与障碍物地图;②决策层在感知结果上做任务分解与规划,包括全局路径规划(A*/RRT)与局部行为决策(避障、跟车);③控制层将决策出的轨迹转化为电机/舵机指令,常用 PID、MPC、轨迹跟踪控制器执行;④执行后环境改变,再次被感知,形成反馈闭环。该闭环的实时性与鲁棒性决定系统成败——感知延迟、决策错误或控制抖振都会累积,因此各环节需低时延且相互解耦校准。同济智能车实验室正是按此闭环训练学生的系统集成能力。
Reference Answer (EN)
The loop: (1) Perception—sensors (camera, LiDAR, IMU) plus SLAM/detection produce pose and obstacle map; (2) Decision—global planning (A*/RRT) and local behavior (avoidance, following); (3) Control—convert trajectory into actuator commands via PID/MPC; (4) act, observe changed environment, loop again. Latency and robustness of each stage determine system success.
备考提示:建议画一张简图辅助表达:传感器→SLAM→规划→控制器→执行器→环境→回授。能画图说明的同学更受青睐。
Q4简述强化学习中的"策略(Policy)"是什么,如何表示?
EN: What is a policy in RL, and how is it represented?
参考回答(中文)
策略 π(a|s) 是智能体的"行为规则",定义在给定的状态 s 下选择动作 a 的概率分布(随机策略)或确定性映射(确定性策略)。它可由查表(小状态空间)、线性函数或深度神经网络参数化——后者即"深度强化学习",例如用 CNN 把图像状态映射为动作概率。智能体的目标就是学习一个策略,使其长期累积奖励期望最大。在机器人任务中,策略常把"图像/本体感受"直接映射到"关节速度/转向角",是端到端具身控制的核心组件。
Reference Answer (EN)
A policy π(a|s) is the agent's behavior rule: a probability (stochastic) or deterministic mapping from state s to action a. It can be tabular, linear, or—most commonly—a deep neural network (deep RL). The goal is to learn a policy maximizing expected cumulative reward. In robotics, policies often map images/proprioception directly to joint velocities or steering angles.
备考提示:区分"策略"与"价值函数"是常考点,下一题接着问价值,注意衔接。
Q5价值函数(Value Function)与动作价值 Q 函数有何区别?它们有什么用?
EN: How do value functions and Q-functions differ, and what are they for?
参考回答(中文)
状态价值函数 V(s) 表示"从状态 s 出发、遵循策略 π 所能获得的未来累积奖励期望",衡量状态的好坏;动作价值 Q(s,a) 表示"在状态 s 采取动作 a 后、再遵循策略 π 的期望回报",衡量"状态-动作对"的好坏。区别在于 Q 额外条件于一个具体动作。二者用途:一是评估策略优劣,二是在决策时通过"选 Q 值最大的动作"来隐式推导出策略(如 Q-learning、DQN)。在机器人控制中,价值函数帮助智能体判断"当前局面是否安全/有利",避免短视动作。
Reference Answer (EN)
V(s) is the expected return from state s under policy π; Q(s,a) is the expected return after taking action a in s then following π. Q conditions on a specific action. They evaluate policy quality and support action selection (choose max-Q), e.g., Q-learning/DQN. In robotics, value functions help judge whether a state is safe/advantageous.
备考提示:Bellman 方程是价值函数的数学核心,若被追问可写出 V(s)=E[R+γV(s')],展示功底。
Q6奖励函数(Reward)设计为什么困难?举一个机器人任务中的例子。
EN: Why is reward design hard? Give a robotics example.
参考回答(中文)
奖励设计难在"稀疏性"与"错位性":稀疏奖励(只有任务完成才给 +1)导致探索效率极低;而设计过密的奖励又易产生"奖励黑客"——智能体钻指标空子却没真正完成任务。例如让机械臂抓取物体,若只以"末端到达目标点"为奖励,机械臂可能用碰撞、推倒的方式"作弊";若每步给"靠近目标"的密集奖励,又可能卡在局部(贴着物体不抓)。常用解法是课程学习、奖励塑形(reward shaping)、或改用目标条件强化学习(HER)。这正体现具身任务"目标定义即难点"的特点。
Reference Answer (EN)
Hard due to sparsity (only terminal +1 → poor exploration) and misalignment (dense rewards cause reward hacking). Example: a grasping arm rewarded only for reaching the target may cheat by pushing; a dense "distance" reward may stall locally. Solutions: curriculum learning, reward shaping, Hindsight Experience Replay (HER).
备考提示:能举出"奖励黑客"实例,说明你真正跑过 RL 实验,比空谈理论更有说服力。
Q7专业重点专业重点请解释 PPO 算法为什么在机器人强化学习中常用,它的核心思想是什么?
EN: Why is PPO popular in robotic RL, and what is its core idea?
参考回答(中文)
PPO(Proximal Policy Optimization)是一种策略梯度算法,核心思想是"限制每次策略更新的幅度,避免训练崩溃"。它通过裁剪(clipped)概率比 r(θ)=π_θ(a|s)/π_old(a|s) 的信赖域式目标,使新策略不会偏离旧策略太远,从而在"稳定"与"样本效率"间取得平衡。相比原始策略梯度易因步长过大而发散、相比 TRPO 实现更简单,PPO 成为机器人仿真训练(如训练四足行走、机械臂操作)的默认选择。其缺点是仍属 on-policy、样本利用率偏低,因此在真实机器人上常需配合 Sim2Real 先在仿真中预训练。
Reference Answer (EN)
PPO is a policy-gradient method that limits per-update change via a clipped surrogate objective on the probability ratio r(θ). This prevents destructive large updates, balancing stability and sample efficiency. Simpler than TRPO and more stable than vanilla PG, it is the de-facto choice for simulated robotic training (locomotion, manipulation). Its on-policy nature means lower sample efficiency, so Sim2Real pretraining is common.
备考提示:可对比 SAC(off-policy、样本更高效)与 PPO,显示你对算法选型有判断,而非只会背名词。
Q8A*、Dijkstra 与 RRT 在路径规划中各有什么特点?如何选用?
EN: Compare A*, Dijkstra, and RRT for path planning; how to choose?
参考回答(中文)
Dijkstra 是在图/栅格上求单源最短路径的完备算法,等权扩展、保证最优但效率低;A* 在 Dijkstra 基础上引入启发函数 h(s)(如欧氏距离)引导搜索方向,显著提速且仍保证最优(当 h 可采纳时);RRT(快速随机树)则面向高维连续空间与非完整约束(如无人机、机械臂),通过随机采样逐步扩展树,概率完备但不保证最优、路径常需后处理平滑。选用原则:低维栅格/已知地图用 A*;高维、复杂约束、实时性优先用 RRT/RRT*;RRT* 可渐近最优。
Reference Answer (EN)
Dijkstra: complete, optimal, slow (uniform expansion). A*: adds admissible heuristic h(s) to guide search—faster, optimal. RRT: samples randomly in high-dimensional/nonholonomic spaces (drones, arms), probabilistically complete, not optimal unless RRT* (asymptotically optimal), path needs smoothing. Choose A* for low-D grids; RRT* for high-D, constrained, real-time cases.
备考提示:同济智能车竞赛常用栅格地图 A* 或 Hybrid A*,能点出"Hybrid A* 兼容车辆运动学约束"会很亮眼。
Q9A* 算法中的启发函数 h(n) 需要满足什么条件才能保证最优?解释"可采纳性"。
EN: What condition must A*'s heuristic satisfy for optimality? Explain admissibility.
参考回答(中文)
A* 在 h(n) 满足"可采纳性(admissible)"且一致(consistent)时保证找到最优路径。可采纳指 h(n) 对任意节点 n 都不高估到达目标的真实最低代价,即 h(n)≤h*(n)。由于估计从不低于真实值,A* 不会错误地剪掉最优路径;若再满足一致性(三角不等式 h(n)≤c(n,n')+h(n')),则每个节点只需扩展一次、效率更高。实践中常用欧氏距离或曼哈顿距离作启发,它们均是可采纳的。要注意可采纳保证最优但可能扩展较多节点,故常用加权 A* 以精度换速度。
Reference Answer (EN)
A* is optimal when h(n) is admissible and consistent. Admissible means h(n) never overestimates the true cost-to-go: h(n) ≤ h*(n). This prevents pruning the optimal path. Consistency (h(n) ≤ c + h(n')) adds efficiency (each node expanded once). Euclidean/Manhattan distance are admissible heuristics. Weighted A* trades optimality for speed.
备考提示:"可采纳"与"一致"常被混淆,能说清二者关系(一致⇒可采纳,反之不成立)体现严谨性。
Q10RRT 为什么适合高维空间?它有什么缺点,如何改进?
EN: Why does RRT suit high-dimensional spaces? Its drawbacks and improvements?
参考回答(中文)
RRT 适合高维是因为它不显式离散化整个空间,而是随机采样并连接最近节点,从而"绕过"维度灾难,对机械臂(6+维构型空间)、无人机很有效。缺点:①仅概率完备、不保证最优;②路径曲折、贴近障碍;③随机性强、可复现性差。改进有:RRT* 通过"重连"渐近最优;Informed RRT* 在找到首解后限定椭圆采样区域加速收敛;BIT* 结合图搜索与采样。工程上通常再叠加路径平滑(B样条/最短路径简化)以满足机器人动力学约束。
Reference Answer (EN)
RRT avoids the curse of dimensionality by random sampling and connecting to the nearest node, not by discretizing space—ideal for high-D arms/drones. Drawbacks: probabilistic completeness only, non-optimal, zig-zag paths, poor reproducibility. Improvements: RRT* (rewiring, asymptotically optimal), Informed RRT* (ellipsoidal focusing), BIT*; plus post-smoothing (B-spline) for dynamics.
备考提示:若你做过无人机航点规划,可说明"RRT* + 贝塞尔平滑"的实际用法,体现工程经验。
Q11专业重点专业重点请说明 SLAM 的基本原理,并区分前端(前端里程计)与后端。
EN: Explain SLAM basics and distinguish front-end from back-end.
参考回答(中文)
SLAM 解决"机器人在未知环境中一边定位自己、一边构建环境地图"的鸡生蛋问题。其流程分两部分:前端(里程计)负责从传感器数据(视觉特征、激光扫描)做数据关联与相对位姿估计,如视觉 SLAM 中的特征提取-匹配-求 pose,或激光的帧间 ICP 配准,输出"增量运动";后端负责全局优化,把前端累积的带噪声约束通过图优化/Bundle Adjustment/滤波(EKF)联合优化,消除漂移、得到一致地图与轨迹。常见方案有视觉 SLAM(ORB-SLAM)、激光 SLAM(Cartographer)及多传感器融合(LIO)。SLAM 是同济无人系统导航定位的底层支柱。
Reference Answer (EN)
SLAM jointly estimates robot pose and builds a map in an unknown environment. Front-end (odometry): from sensor data (visual features, LiDAR scans) estimate relative pose via data association—feature matching or ICP—producing incremental motion. Back-end: global optimization (graph optimization, BA, EKF) fuses noisy constraints to remove drift and yield a consistent map/trajectory. Examples: ORB-SLAM, Cartographer, LIO.
备考提示:"闭环检测(loop closure)"是后端纠漂的关键,被追问时务必能讲清它如何触发全局优化。
Q12激光 SLAM 与视觉 SLAM 各有什么优劣?什么场景用哪种?
EN: Compare LiDAR SLAM vs visual SLAM: pros, cons, and use cases.
参考回答(中文)
激光 SLAM(如 Cartographer、LOAM)直接测距、对光照与纹理不敏感、建图精度高、鲁棒性强,缺点是无纹理退化场景(长走廊)易失效、成本较高、缺乏语义。视觉 SLAM(如 ORB-SLAM、VINS)传感器便宜、信息丰富、可顺带做语义/目标识别,但受光照、动态物体、弱纹理影响大,单目还有尺度不确定问题(需 IMU 或双目)。选用:室内结构化、室外大场景、追求稳定用激光;成本敏感、需语义理解、轻量平台用视觉/视觉-惯性。实际系统中常多传感器融合以互补。
Reference Answer (EN)
LiDAR SLAM: direct range, lighting/texture-invariant, accurate, robust; but fails in degenerate corridors, costlier, no semantics. Visual SLAM: cheap, rich semantics, but sensitive to light/dynamics, monocular lacks scale (needs IMU/stereo). Use LiDAR for robust indoor/outdoor mapping; visual/VINS for low-cost, semantic, lightweight platforms. Fusions are common.
备考提示:能提到"退化场景"和"多传感器融合"说明你理解工程落地痛点,而非只会背定义。
Q13什么是视觉伺服(Visual Servoing)?它和"先定位再规划"的范式有何不同?
EN: What is visual servoing, and how does it differ from "localize-then-plan"?
参考回答(中文)
视觉伺服是直接以图像特征误差作为反馈信号来控制机器人运动的闭环方法,分为 IBVS(基于图像)和 PBVS(基于位置)。与传统"SLAM 定位→全局规划→跟踪"的模块化范式不同,视觉伺服省略中间建图/定位步骤,直接用"当前图像特征与目标特征的偏差"驱动控制器(如把目标物保持在图像中心)。优点是反应快、对模型误差鲁棒、适合精细操作(抓取、对准);缺点是视场受限、易丢失特征。它体现"感知即控制"的端到端思想,与强化学习端到端控制有相似之处。
Reference Answer (EN)
Visual servoing uses image-feature error directly as feedback to control motion (IBVS/PBVS). Unlike modular "localize→plan→track", it skips mapping/localization and drives the controller straight from the gap between current and desired image features (e.g., keep target centered). Pros: fast, robust to model error, good for fine manipulation. Cons: limited field of view, feature loss.
备考提示:对比"模块化 vs 端到端"是高频论点,视觉伺服是绝佳例证,可记一句:"它把感知直接闭环到控制。"
Q14什么是模仿学习?行为克隆(Behavior Cloning)有什么局限?
EN: What is imitation learning? What are the limits of behavior cloning?
参考回答(中文)
模仿学习是从专家示范中习得策略的范式,避免手工设计奖励。行为克隆是其最简单形式——把示范看成 (状态, 动作) 数据集,用监督学习直接拟合 π(a|s),如同"学徒照着师傅做"。其局限有三:一是复合误差(covariate shift)——执行中小偏离会进入训练时未见过状态,误差累积致"越偏越远";二是分布不匹配,专家不会示范"犯错后如何恢复";三是数据覆盖不足。改进办法包括 DAgger(让专家在智能体访问到的状态上补标)、逆强化学习(IRL)推断奖励再学习。同济无人机避障常先用行为克隆快速出原型。
Reference Answer (EN)
Imitation learning learns from expert demonstrations to avoid hand-crafted rewards. Behavior cloning is its simplest form: supervised learning on (state, action) pairs. Limits: (1) compounding error / covariate shift—small deviations enter unseen states, errors accumulate; (2) experts never demonstrate recovery from mistakes; (3) poor coverage. Fixes: DAgger (query expert on visited states), inverse RL.
备考提示:务必能讲清"复合误差"与 DAgger 的对应关系,这是区分背概念与真理解的关键。
Q15什么是 Sim2Real?常用哪些技术缩小仿真与现实的差距?
EN: What is Sim2Real, and what techniques bridge the sim-real gap?
参考回答(中文)
Sim2Real 指将在仿真中训练的模型/策略迁移到真实机器人的过程,核心障碍是"域差距"——仿真物理、渲染、传感器噪声与现实不一致,导致策略在真机上失效。常用技术:①域随机化(Domain Randomization),在仿真中随机化摩擦、光照、质量等参数,迫使策略学不变特征;②系统辨识,先拟合真实动力学再注入仿真;③域自适应/对抗,把仿真图像映射到真实风格;④渐进式迁移与真实数据微调。这是同济"先仿真训练、再真机部署"具身项目的标准流程,能显著降低真机试错成本。
Reference Answer (EN)
Sim2Real transfers simulation-trained policies to real robots; the obstacle is the domain gap (physics, rendering, sensor noise mismatch). Techniques: (1) Domain Randomization—randomize friction, lighting, mass so the policy learns invariant features; (2) system identification to fit real dynamics; (3) domain adaptation/adversarial style transfer; (4) progressive transfer and real-data fine-tuning.
备考提示:"域随机化"几乎必考,需能解释为何"随机得越狠、迁移越稳"(在分布内最差情况也鲁棒)。
Q16专业重点专业重点请说明正运动学与逆运动学,并解释逆解为什么通常更难。
EN: Explain forward and inverse kinematics, and why inverse kinematics is harder.
参考回答(中文)
正运动学(FK)是在已知各关节角的情况下,通过连杆变换(如 Denavit-Hartenberg 参数)计算机械臂末端位姿——这是唯一确定的映射。逆运动学(IK)则反过来:给定期望末端位姿求所需关节角。IK 更难的原因:一是多解性(同一末端位姿可能有多组关节解,需选"最优");二是奇异性——在某些构型下雅可比矩阵降秩,微小末端位移对应极大/无解关节变化;三是约束(关节限位、避障)使可行解受限。工程上常用数值法(雅可比伪逆、阻尼最小二乘)迭代求解,而非解析解。
Reference Answer (EN)
Forward kinematics (FK): given joint angles, compute end-effector pose via link transforms (DH parameters)—a unique mapping. Inverse kinematics (IK): given a desired pose, find joint angles. IK is harder because of (1) multiple solutions; (2) singularities where the Jacobian loses rank; (3) constraints (joint limits, collision). Numerical methods (Jacobian pseudoinverse, DLS) are typically used.
备考提示:被追问时画出"雅可比矩阵连接关节速度到末端速度(v=J·q̇)",能展现扎实基础。
Q17雅可比矩阵在机器人学中起什么作用?举一个应用。
EN: What role does the Jacobian play in robotics? Give an application.
参考回答(中文)
雅可比矩阵 J 描述了"关节空间速度"与"末端操作空间速度"之间的线性映射:ẋ = J(q)·q̇,是连接运动学与控制的桥梁。它的作用包括:①逆运动学数值求解(用 J 的伪逆迭代逼近目标位姿);②奇异性分析(det(JJᵀ) 小即接近奇异);③力/力矩映射(静力下 τ = JᵀF);④冗余机械臂的零空间优化(避开奇异与限位)。例如机械臂用阻尼最小二乘(JᵀJ+λI)求逆,可在接近奇异时保持稳定跟踪。
Reference Answer (EN)
The Jacobian J maps joint velocities to end-effector velocities: ẋ = J(q)q̇. Uses: (1) numerical IK via pseudoinverse; (2) singularity analysis (small det(JJᵀ)); (3) force mapping τ = JᵀF; (4) null-space optimization for redundant arms. Example: damped least squares (JᵀJ+λI)⁻¹ keeps tracking stable near singularities.
备考提示:把雅可比同时联系 IK 与力控,能展示你理解它的"中枢"地位,而非孤立记忆。
Q18PID 控制器三个参数分别代表什么?机器人控制中如何调参?
EN: What do the three PID terms mean, and how do you tune them in robotics?
参考回答(中文)
PID 由比例 P、积分 I、微分 D 三项构成:P 按当前误差成比例出力,响应快但有余差;I 累积历史误差消除稳态偏差,但过强易振荡;D 依据误差变化率提供阻尼,抑制超调。调参常用"先 P 后 D 再 I"的工程法:先加大 P 至系统开始振荡,回退到稳定临界值;再加 D 抑制超调;最后加少量 I 消除余差。机器人速度/位置跟踪多用 PI,姿态控制加 D 防抖。现代也常用 MPC 处理带约束的多变量控制。
Reference Answer (EN)
P: proportional to current error (fast but steady-state offset); I: integrates past error (removes offset, risk of oscillation); D: derivative damps overshoot. Tuning rule: increase P until slight oscillation, back off, add D to reduce overshoot, then small I for offset. Robotics: PI for speed/position, add D for attitude. MPC handles constrained multivariable control.
备考提示:能说明"积分饱和(windup)"及其限幅处理,是控制实操经验的加分项。
Q19专业重点专业重点多机器人协同控制有哪些典型范式?各有什么优劣?
EN: What are typical multi-robot coordination paradigms and their trade-offs?
参考回答(中文)
多机器人协同主要有三类范式:①集中式——由中心节点统一规划,全局最优但单点故障、通信负担大、可扩展性差;②分布式/去中心化——各机器人基于本地与邻居信息自主决策,鲁棒可扩展,但难保证全局最优,需一致性算法(如 consensus)协调;③分层式——群体分簇、簇头协调,折中二者。协同任务常需解决一致性(consensus)、编队控制(leader-follower、虚拟结构)、任务分配(拍卖/市场机制)与防碰撞。同济"无人系统协同控制"前沿中心正是聚焦该类问题,如无人机-车协同侦察。
Reference Answer (EN)
Three paradigms: (1) Centralized—optimal but single-point failure, heavy comms, poor scalability; (2) Distributed—robust, scalable via consensus, but no global optimality guarantee; (3) Hierarchical—clusters with leaders, a trade-off. Key problems: consensus, formation control (leader-follower, virtual structure), task allocation, collision avoidance.
备考提示:能点出"一致性算法(consensus)"与"通信拓扑"的关系,紧扣同济无人系统协同特色。
Q20无人机自主避障通常需要哪些模块协同?请描述一个简化系统架构。
EN: What modules does autonomous drone obstacle avoidance need? Describe a simplified architecture.
参考回答(中文)
一个简化的无人机避障系统包含:①感知——双目/深度相机或激光雷达获取障碍物,IMU 提供姿态;②状态估计——视觉惯性里程计(VIO)或 SLAM 给出自身位姿与局部地图;③局部规划——在动态障碍下实时生成无碰撞轨迹(如基于 B 样条 + 梯度下降的局部规划器、或模型预测控制);④控制——将轨迹转为姿态/推力指令(PID/几何控制);⑤决策层处理任务与应急(低电量返航、失联悬停)。各模块须低时延(通常 >20Hz)协同;同济无人机方向重点即此类"感知-规划-控制"一体化设计。
Reference Answer (EN)
Modules: (1) Perception—stereo/depth camera or LiDAR, IMU; (2) State estimation—VIO/SLAM for pose and local map; (3) Local planning—real-time collision-free trajectory (B-spline + gradient, or MPC); (4) Control—attitude/thrust via PID/geometric control; (5) Decision—mission logic, failsafe (RTL, hover). All must run at low latency (>20 Hz) and tightly integrate.
备考提示:"VIO"与"几何控制"是无人机专有术语,能自然带出说明你对平台有真实了解。
Q21专业重点专业重点什么是 VLA(视觉-语言-动作)模型?它如何"用大模型赋能机器人"?
EN: What is a VLA (vision-language-action) model, and how does it empower robots with LLMs?
参考回答(中文)
VLA 是具身智能的新范式:把视觉观测、语言指令与机器人动作统一到一个大模型里,直接输出可执行动作(或动作 token)。典型如 RT-1/RT-2、OpenVLA——以多模态大模型为骨干,将"看到的图像+听懂的指令"映射为"该做的动作序列",借助预训练大模型的语义泛化能力,使机器人能零样本理解新物体与新指令。它用大模型赋能机器人的方式有三:提供常识与语义理解、跨任务迁移、自然语言交互。局限是需大量真实示范、实时性与安全性仍需工程保障。这与同济对接上海人工智能实验室的具身方向高度相关。
Reference Answer (EN)
VLA unifies vision, language, and action in one model that outputs executable actions (or action tokens), e.g., RT-2, OpenVLA. Using a pretrained multimodal LLM backbone, it maps image + instruction to action sequences, enabling zero-shot generalization to new objects/commands. LLMs empower robots via semantic understanding, cross-task transfer, and natural-language interaction. Limits: needs many real demonstrations; real-time safety remains engineering work.
备考提示:能区分"VLA 端到端出动作"与"LLM 只做高层规划+底层控制器执行"两种路线,体现前沿视野。
Q22大模型(LLM)在机器人系统中一般承担什么角色?会直接控制电机吗?
EN: What role do LLMs play in robotic systems? Do they directly control motors?
参考回答(中文)
大模型在机器人中通常扮演"高层语义大脑"而非底层执行器:①任务理解与分解——把"把桌子收拾干净"拆成抓取、移动、放置等子任务;②代码/技能生成——输出控制 API 调用或 Python 脚本(如 code-as-policy);③常识与故障解释。它一般不直接输出电机力矩,因为大模型时延高、缺乏精确数值与实时反馈,底层仍由运动规划与传统控制器闭环执行。少数 VLA 模型会直接出动作,但主流方案是"LLM 规划 + 专用控制器执行"的分层架构,兼顾智能与稳定。
Reference Answer (EN)
LLMs usually act as a high-level semantic brain, not low-level actuators: (1) task understanding/decomposition; (2) code/skill generation (code-as-policy); (3) commonsense and failure explanation. They generally do not directly output motor torques—too slow and imprecise. Low-level execution stays with planners and controllers. Mainstream: "LLM plans, dedicated controller executes."
备考提示:"分层架构"是关键词,强调"大模型负责想、控制器负责做",可避免被质疑安全性。
Q23如果你参与过智能车竞赛,会如何向面试官介绍你的项目与技术栈?
EN: How would you present a smart-car competition project to an interviewer?
参考回答(中文)
我会用"问题—方法—结果"的结构:先说明任务(如循迹、锥桶识别、超车),再讲技术栈——感知用相机/激光做车道线与障碍物检测,定位用 SLAM 或 GNSS/IMU 融合,规划用 Hybrid A* 或 RRT 出全局路径、纯跟踪(Pure Pursuit)做轨迹跟踪,控制用 PID/滑模。重点突出一个我解决的具体难题(如动态障碍误检、弯道抖动),并说明指标提升。最后点明该项目让我理解了"感知-决策-控制"闭环与系统工程,与同济智能车竞赛实验室的培养方向一致,体现我能把算法跑在真车上。
Reference Answer (EN)
Use problem–method–result: state the task (tracking, cone detection, overtaking); describe the stack—perception (camera/LiDAR), localization (SLAM or GNSS/IMU), planning (Hybrid A*/RRT + Pure Pursuit), control (PID/sliding mode). Highlight one concrete problem solved and the metric gain. Conclude by linking it to the perceive-decide-control loop and Tongji's smart-car lab.
备考提示:评委最怕"只挂名没干活",务必准备一个你能讲清细节的技术点(哪怕是调参经历)。
Q24专业重点专业重点针对"无人机自主避障",如果让你从零设计一个方案,你会怎么分层?
EN: Design a drone autonomous-avoidance system from scratch—how would you layer it?
参考回答(中文)
我会按四层设计:①感知层——用深度相机/激光雷达测距,IMU 测姿态,做传感器标定与时间同步;②建图层——以 VIO/SLAM 输出里程计,构建占据栅格或 ESDF 局部地图;③规划层——全局用 A*/RRT* 出粗路径,局部用基于 ESDF 的梯度规划或 MPC 实时绕障,并对动态障碍做速度预测;④控制层——几何控制器把期望轨迹转为姿态与推力,加安全监视(看门狗)在异常时悬停。关键保证各层频率匹配(感知 30Hz、控制 200Hz+)与故障降级,体现同济无人系统"安全优先"的工程观。
Reference Answer (EN)
Four layers: (1) Perception—depth/LiDAR + IMU, calibrated and time-synced; (2) Mapping—VIO/SLAM odometry, occupancy grid or ESDF local map; (3) Planning—global A*/RRT*, local ESDF-gradient or MPC avoidance with dynamic-obstacle velocity prediction; (4) Control—geometric controller to attitude/thrust, plus a watchdog for safe hover on anomaly. Match frequencies (perception 30Hz, control 200Hz+) with graceful degradation.
备考提示:主动提到"ESDF(欧氏符号距离场)"和"故障降级",比泛泛而谈更显专业与安全意识。
Q25强化学习直接用在真实机器人上有哪些困难?业界如何应对?
EN: What difficulties arise applying RL directly on real robots, and how is it addressed?
参考回答(中文)
直接上真机 RL 的困难:①样本效率低——真机采集慢且易损耗,on-policy 方法需海量交互;②安全——探索中乱动作会撞坏机器或伤人或破坏环境;③重置困难——任务结束后难自动复位;④现实噪声与非平稳。业界对策:先在仿真用 PPO/SAC 预训练,再 Sim2Real 迁移(域随机化);用安全 RL(约束策略、屏障函数);或离线 RL(从已有数据集学,不在线探索);以及现实世界少步微调。同济实验室多采用"仿真为主、真机微调"的稳妥路线。
Reference Answer (EN)
Difficulties: (1) low sample efficiency on slow/damaging real hardware; (2) safety during exploration; (3) hard auto-reset; (4) real noise/non-stationarity. Responses: pretrain in sim (PPO/SAC) then Sim2Real with domain randomization; safe RL (constraints, barrier functions); offline RL from datasets; few-shot real fine-tuning. Tongji labs favor "sim-first, real-finetune."
备考提示:能说出"离线 RL"和"安全屏障函数",说明你关注前沿而非只懂教科书 RL。
Q26动态障碍物环境下的路径规划与静态环境有何本质不同?如何应对?
EN: How does planning with dynamic obstacles differ from static, and how to handle it?
参考回答(中文)
本质不同在于:静态规划只需"无碰撞的几何路径",而动态环境要求"时空上都安全",即不仅要避开位置,还要避开障碍物未来会出现的位置。应对方法:①预测——用卡尔曼滤波/学习型模型估计障碍速度并外推;②时变代价图——把时间维加入规划(如时空 A*、D* Lite 增量重规划);③速度障碍法(Velocity Obstacle)/模型预测控制,直接在速度空间求解;④局部重规划高频刷新。多机器人场景还需协作预测彼此轨迹以防死锁。同济无人系统协同常以此类方法来做动态编队。
Reference Answer (EN)
Key difference: static planning needs a collision-free geometric path; dynamic needs spatiotemporal safety—avoiding where obstacles will be. Methods: (1) prediction (Kalman/learned velocity extrapolation); (2) time-varying cost maps / spatio-temporal A*, D* Lite incremental replanning; (3) Velocity Obstacles / MPC in velocity space; (4) high-rate local replanning. Multi-robot needs mutual trajectory prediction to avoid deadlock.
备考提示:"时空规划"和"速度障碍法"是动态避障的两把钥匙,能讲清其一即可。
Q27机械臂抓取任务中,如何表示和计算"抓取姿态"?涉及哪些感知要素?
EN: How is a grasp pose represented and computed? What perception is involved?
参考回答(中文)
抓取姿态通常用 6-DoF 位姿(位置 + 姿态,常以四元数/旋转矩阵表示)描述夹爪相对物体的目标位形,常以"抓取矩形"或 SE(3) 变换表达。计算流程:①感知——用 RGB-D/点云做物体检测与分割,估计位姿(如基于 ICP 或 PoseCNN);②候选生成——采样若干可行抓取(如平行夹爪的两指接触点对,满足摩擦锥与夹持宽度约束);③评分——用几何指标(摩擦、夹持力、碰撞)或学习型抓取质量网络(GQ-CNN)打分选最优;④执行——逆运动学求关节角、力控闭合。强调"感知到位姿、规划到无碰、控制到力控"的闭环。
Reference Answer (EN)
A grasp pose is a 6-DoF SE(3) transform (position + orientation) of the gripper relative to the object, often as a grasp rectangle. Pipeline: (1) perception—RGB-D/point-cloud detection, segmentation, pose estimation (ICP, PoseCNN); (2) candidate generation—sample antipodal contacts satisfying the friction cone and width; (3) scoring—geometric metrics or learned GQ-CNN; (4) execution—IK + force-controlled closure.
备考提示:"摩擦锥(friction cone)"和"对映抓取(antipodal)"是抓取力学核心,能带出即专业。
Q28卡尔曼滤波在机器人状态估计中做什么?它基于什么假设?
EN: What does the Kalman filter do in robot state estimation, and what assumptions does it make?
参考回答(中文)
卡尔曼滤波(KF)是一种递推的最优状态估计算法,融合"带噪声的过程模型预测"与"带噪声的观测",输出最小均方误差的状态估计,广泛用于定位、SLAM 前端、传感器融合(如 IMU+GPS)。它基于两个假设:①系统为线性、②过程与观测噪声均为零均值高斯且相互独立。当系统非线性时改用扩展卡尔曼(EKF,局部线性化)或无迹卡尔曼(UKF);多传感器、大状态时常用因子图优化替代。KF 是机器人"在不确定中做可靠估计"的基础工具。
Reference Answer (EN)
The Kalman filter recursively fuses a noisy process model and noisy measurements into a minimum-variance state estimate—widely used in localization, SLAM front-ends, and sensor fusion (IMU+GPS). Assumptions: (1) linear system; (2) zero-mean Gaussian, independent process/measurement noise. For nonlinear systems use EKF (linearization) or UKF; large/multi-sensor problems often use factor-graph optimization.
备考提示:能说出"EKF 是 KF 的非线性扩展、但线性化会引入误差",显示你理解假设与局限。
Q29具身智能中"端到端"与"模块化(pipeline)"两种系统架构各有什么取舍?
EN: Compare end-to-end vs modular (pipeline) architectures in embodied AI.
参考回答(中文)
模块化把系统拆为感知、定位、规划、控制独立模块,优点是可解释、易调试、各模块可用成熟方法(如 SLAM、A*),缺点是误差逐级传递、接口手工设计、难全局优化;端到端用单一网络从原始输入直接输出动作(如视觉→动作),优点是信息无损、可联合优化、潜力上限高,缺点是数据需求大、黑箱难调试、不安全时后果严重。实践中常折中:感知端到端、控制仍模块化,或 VLA 做高层、传统控制器做底层。同济工程智能强调"可靠落地",因此模块化仍是主流。
Reference Answer (EN)
Modular splits perception/localization/planning/control—interpretable, debuggable, uses mature methods, but error accumulates across stages and interfaces are hand-designed. End-to-end maps raw input straight to action—lossless, jointly optimized, higher ceiling, but data-hungry, black-box, risky. Practice hybrid: end-to-end perception + modular control, or VLA high-level + traditional low-level. Tongji's AI4E favors reliable modular deployment.
备考提示:回答架构题务必给"取舍"而非站队,评委看重工程权衡思维。
Q30你如何看待具身智能与机器人学的未来 3–5 年趋势?同济在该方向能给你什么?
EN: What are the next 3–5 year trends in embodied AI, and what can Tongji offer you?
参考回答(中文)
趋势有三:①基础模型化——VLA 与"机器人基础模型"将像大模型一样成为通用底座;②仿真规模化——以仿真工厂+Sim2Real 大规模产出技能;③软硬协同——更便宜灵巧的执行器与触觉传感。挑战仍在数据、安全与泛化。对同济,我认为价值在:依托电子与信息工程学院与"自主智能无人系统前沿科学中心"的真实平台(智能车、无人机、VEX、协同控制实验室),以及和上海人工智能实验室的具身合作、项目制+导师制培养,能让我"在真机上做科研",这正是具身方向最稀缺的训练环境。
Reference Answer (EN)
Trends: (1) foundation models—VLA and robot foundation models as general backbones; (2) scaled simulation—sim factories + Sim2Real for mass skill generation; (3) hardware-software co-design—cheaper dexterous actuators and tactile sensing. Challenges remain in data, safety, generalization. Tongji offers real platforms (smart cars, drones, VEX, coordination labs), the frontier science center, Shanghai AI Lab collaboration, and project/tutor dual-track training—letting me do research "on real robots."
备考提示:结尾题要落回"我为什么选同济、同济能给我什么",把学校特色自然嵌进个人志向,印象分最高。