未知環境中機器人巡航問題之研究 王銀添 副教授 機器人實驗室 淡江大學機械與機電工程學系 Applied Electronics Technology, NTNU 1/48 目錄 • • • • • 機器人巡航 感測器輔助機器人執行巡航任務 可能遭遇的問題與解決方案 不確定性(uncertainty)現象 機率式狀態估測方法 – 貝氏規則、Kalman Filter、Particle filter – 同時定位、建圖、物件追蹤之高維度非線性系統 • 同時定位、建圖、物件追蹤實測範例 • 結論與未來之研究議題 Applied Electronics Technology, NTNU 2/48 機器人巡航 • 在未知的環境中巡航時,機器人想知道 – 自己在哪裡? – 環境是什樣的長相? – 是否有移動的障礙物? Example: Dead reckoning (deduced reckoning) x p y Applied Electronics Technology, NTNU vx p p v y t 3/48 Calibration of Errors for Robot with Odometry (Borenstein [1992]) • The unidirectional square path experiment Applied Electronics Technology, NTNU 4/48 Calibration of Errors for Robot with Odometry (Borenstein [1992]) • The bi-directional square path experiment Applied Electronics Technology, NTNU 5/48 感測器輔助巡航 • 在未知的環境中巡航時,機器人必須依賴自身的 移動與(外部、多個)感測器對環境特徵的感測, 執行以下任務: – 自我定位(self-localization)任務 – 環境地圖建構(mapping)任務 – 移動物體偵測與追蹤(detection and tracking of moving objects)任務 Applied Electronics Technology, NTNU 6/48 機器人感測(Robot Perception)系統 Sensor Classification • Proprioceptive sensors – measure values internally to the system (robot) – e.g. motor speed, wheel load, heading of the robot, battery status • Exteroceptive sensors – information from the robots environment – distances to objects, intensity of the ambient light, unique features • Passive sensors – energy coming for the environment – e.g. temperature probe, microphones, and CCD or CMOS camera. • Active sensors – emit their proper energy and measure the environmental reaction – better performance, but some influence on environment – e.g. wheel quadrature encoders, ultrasonic sensors, and laser rangefinders. Applied Electronics Technology, NTNU 7/48 [Siegwart and Nourbakhsh 2004] Sensor Classification (1) Applied Electronics Technology, NTNU 8/48 Sensor Classification (2) Applied Electronics Technology, NTNU 9/48 Laser Range Finder (LRF) ( x )2 ( x )2 v x x y y r r g (x) y xy tan1( ) v x x x Applied Electronics Technology, NTNU 10/48 Mapping Using LRF Applied Electronics Technology, NTNU 11/48 Inertial Measurement Unit (IMU) • A unit has 3-axis gyroscope (pitch, roll, yaw ) and 3-axis accelerometer. • Localization using IMU • A unit with a tri-axis accelerometer, tri-axis magnetometer and a tri-axis gyro Applied Electronics Technology, NTNU 12/48 Vision Sensors image plane Iy camera center Ix (u0 ,v0) hC(Ix ,Iy ) fC zC (0,0,0) C hC(hxC, hC y , hz ) P xC yC {c} zc xc yc hic Image projection model (3D to 2D) hxC u 0 f u C I x hz hC I y y v fv 0 hzC rw Yi w z {w} beacon y x Applied Electronics Technology, NTNU 13/48 Web Camera Based 3D Scanner Applied Electronics Technology, NTNU 14/48 Light Detection and Ranging (LiDAR) Applied Electronics Technology, NTNU 15/48 Airborne LiDAR Applied Electronics Technology, NTNU 16/48 Microsoft Kinect – 3D depth image and RGB color image in 30fps. – Low-cost. (NT$4,550 tax. included) – Software development kit provided by Microsoft. Applied Electronics Technology, NTNU 17/48 Skanect – Real-time Kinect-based 3D Scanner manctl.com Applied Electronics Technology, NTNU 18/48 Mapping Using Kinect Applied Electronics Technology, NTNU 19/48 可能遭遇的問題與解決方案 • 有幾個問題會造成巡航的任務相當棘手,包括 – 感測器的侷限性 – 移動與量測都具有不確定性質(uncertainty) – 定位、建圖、追蹤物體 系統變成高維度與非線性 • 本研究針對以上問題進行探討,考慮的議題包括 – 感測器的選用、移動偵測 – 不確定性現象的描述 – 同時求解定位、建圖、追蹤物體等問題 • 並且以機率理論解決機器人在未知環境中巡航問 題。 Applied Electronics Technology, NTNU 20/48 不確定性(Uncertainty)現象 • The structured errors – The locations of the CW and CCW clusters. • The random errors – The random distribution of errors in the cluster. – Uncertainty in motion. Borenstein [1992] Applied Electronics Technology, NTNU 21/48 Uncertainty in Robot Motion Applied Electronics Technology, NTNU 22/48 Uncertainty in Robotics • 不確定性現象的描述 – 以參數函數描述不確定性,例如高斯常態分佈 N(x;m,s2) m : mean value s : deviation f(x) is the probability density function (pdf) – 以非參數函數描述不確定性,例如蒙地卡羅模擬 • 以機率理論求解具不確定性的機器人巡航問題 Applied Electronics Technology, NTNU 23/48 系統的狀態與量測 • State sequence • xk f xk 1, uk 1, wk 1 x is the state of the system; u is the input; f is a nonlinear function of the state; w is the uncertainty of the state. Measurement sequence z k g xk , vk z is the measurement of the system; g is nonlinear measurement function; v is the uncertainty of the measurement. vx wx x k x k 1 v y t w y w Model-based Uncertainty state transition Applied Electronics Technology, NTNU u0 I x zk I y v 0 hxC fu hzC v x C v hy y fv hzC Projection model Uncertainty 24/48 Basic Probability Theory • Joint probability: P(X=x and Y=y) = P(x,y) • Conditional probability: P(x|y) is the probability of x, given y, P( x, y ) P( x|y) P( y ) P( x , y ) P( x|y) P( y ) P( y, x) P( y|x) P( x) • Theorem of total probability: If yi constitute a partition of the sample space, then for x in the same space P( x) P( x, yi ) yi P( x) P( x | yi )P( yi ) yi p ( x) p ( x | y ) p ( y ) dy • Bayes rule: if x is a quantity that we would like to infer from y, P( y | x) P( x) likelihood prior P( x y ) P( y ) evidence Applied Electronics Technology, NTNU P ( y | x , z ) P( x | z ) P( y | z ) Conditioning Bayes rule on z. P( x | y, z ) 25/48 Probabilistic Generative Laws • The emergence of state xk might be conditioned on all past states, measurements, and controls, p(xk | x0:k 1, z1:k 1, u1:k ) • If the state x is complete then xk-1 is a sufficient statistic of all previous controls and measurements, u1:k-1 and z1:k-1. Only the control uk matters if we know the state xk-1, p(xk | x0:k 1, z1:k 1, u1:k ) p(xk | xk 1, uk ) called state transition probability. • If xk is complete, the measurement probability is also generated by p(zk | x0:k , z1:k 1, u1:k ) p(zk | xk ) The state xk is sufficient to predict the measurement zk. Applied Electronics Technology, NTNU 26/48 Dynamic Bayes Network (DBN) • The temporal generative model is known as hidden Markov model (HMM) or dynamic Bayes network (DBN). – The state at time k is stochastically dependent on the state at time k-1 and the control uk. – The measurement zk depends stochastically on the state at time k. • The dynamic Bayes network that characterizes the evolution of controls, states, and measurements. Applied Electronics Technology, NTNU 27/48 State Estimation Using Bayes’ Rule • From a Bayesian perspective, the state estimation is to recursively calculate some degree of belief in the state xk at time k, given the data z1:k and u1:k, p(xk | z1:k , u1:k ) • Thus, the probability density function (pdf) of state is P ( y | x , z ) P( x | z ) P( x | y, z ) constructed via Bayes’ rule P( y | z ) p(zk | x k ) p(xk | z1:k 1, u1:k ) p(xk | z k , z1:k 1, u1:k ) p(zk | z1:k 1, u1:k ) • The initial pdf p(x0|z0)=p(x0) of the state vector, which is also known as the prior, is available. • z0 is the set of no measurements. • Then, in principle, the pdf p(xk|z1:k,u1:k) may be obtained, recursively, in two stages: prediction and update. Applied Electronics Technology, NTNU 28/48 State Estimation Using Bayes’ Rule Predict and update the state p(xk|z1:k,u1:k) recursively: • The prediction stage involves using the system model to obtain the prior pdf of the state at time k, p(xk | z1:k 1, u1:k ) p(xk | xk 1, u1:k ) p(xk 1| z1:k 1, u1:k 1)d xk 1 Suppose that the required pdf p(xk-1|z1:k-1,u1:k) at time k-1 is available. • At time step k, a measurement zk is used to update the prior (update stage) via Bayes’ rule p(zk | xk ) p(xk | z1:k 1, u1:k ) p(xk | z1:k , u1:k ) p(zk | z1:k 1, u1:k ) where the normalizing constant p(zk |z1:k 1, u1:k ) p(zk |xk )p(xk |z1:k 1, u1:k )d xk Applied Electronics Technology, NTNU 29/48 State Estimation Using Particle Filter • 引用貝氏規則之遞迴預測與更新系統狀態; • 每個粒子代表一個解,在取樣空間中隨機規劃數量L個粒子進行求解 。規劃的粒子數量越多越趨近最佳解。第l個粒子的pdf表示為 p( x[kl ]|z1:k , u1:k ) ~ x[kl ] , w[kl ] 每個粒子l遞迴地依據感測訊息更新取樣空間中的狀態 x[kl ] 之權重值 w[lk ] ,用以顯示狀態在該數值區段的機率。 Particle 2 Given : xk -1, landmarkposition l] dl z[kl ] z[k|k 1 w[kl ] 1 dl 2] z[k|k 1 z[k2] y d2 landmark Model-based z k |k 1 x k |k 1 x Particle 1 At time k-1 z [k1] d1 1] z[k|k 1 At time k Applied Electronics Technology, NTNU 30/48 Procedure of Particle Filter 遞迴地預測與更新系統狀態 • 所有粒子l所存的機器人狀態 x[kl ] 依據運動模型進行移動, 此為粒子的機器人狀態之預測; • 擷取新的視覺感測訊息,並且透過感測模型zk重新分配各 粒子之權重值w[lk ]; • 必要時進行重新取樣(resampling); • 正規化(normalizing)權重值,以及更新機器人的狀態x[kl ]。 Applied Electronics Technology, NTNU 31/48 Particle Filter for Robot Localization Applied Electronics Technology, NTNU 32/48 KF-Based State Estimation Kalman filter (KF) estimator • Adapt the concept of recursive prediction and update estimate process. • Prediction: xk |k 1 A xk 1|k 1 B uk 1 z k |k 1 H xk |k 1 (Linear prediction of states and measurements) • Update: K k Pk |k 1H kT ( H k Pk |k 1H kT Rk )1 Pk|k I Kk Hk Pk|k 1 x [xC m1 m2 mn O1 O2 Ol ]T • State vector of robot (camera) rW rW (vW w k k 1 k 1 vk 1 )t C C C ( w ) t k k 1 k 1 k 1 xCk W v vW wvk 1 k k 1 C C k k 1 wk 1 • State vector of static objects Pk |k 1 Ak Pk 1|k 1 AkT Qk 1 xk|k xk|k 1 Kk z k zk|k 1 Example: KF-based SLAM mik [ X ik Yik Z ik ]T • State vector of moving objects Ojk=[ojk sjk]T Measurement models u0 (Linear update equation for system states) I ix T z Pk E[ek ek ] ik I iy v0 Set Pk / K k 0 tofind K k Applied Electronics Technology, NTNU C hix f C ku C hiz C hiy f C kv C hiz 33/48 Visual Sensors for SLAM • Camera carried by robot • Free-moving camera – The camera is presumed to move at constant velocity (CV); – The acceleration is caused by an impulse noise from the external force. – Velocity noise: wvk ak t w t k k Monocular vision Binocular vision Applied Electronics Technology, NTNU 34/48 淡江機電系機器人實驗室 • 機器人視覺式同時定位、建圖、與移動物體追蹤(visual simultaneous localization, mapping, and moving-object tracking) SLAMMOT SLAM Applied Electronics Technology, NTNU 35/48 Monocular SLAM Applied Electronics Technology, NTNU 36/48 Monocular SLAM Applied Electronics Technology, NTNU 37/48 Monocular People Detection and Tracking Applied Electronics Technology, NTNU 38/48 Binocular SLAM Applied Electronics Technology, NTNU 39/48 Binocular SLAM Applied Electronics Technology, NTNU 40/48 Binocular SLAM Applied Electronics Technology, NTNU 41/48 Differential-drive Mobile Robot [2010] Binocular Vision PC-based Controller Laser Range Finder Sonar Wheel Encoder Mobile Robot Applied Electronics Technology, NTNU 42/48 Visual SLAM of Mobile Robots Applied Electronics Technology, NTNU 43/48 Visual SLAM of Mobile Robots Applied Electronics Technology, NTNU 44/48 Visual SLAM of Mobile Robots Applied Electronics Technology, NTNU 45/48 結論與未來研究議題 • 理論上,同時定位、建圖、與移動物體追蹤的問題已經有 解 • 實現技術方面,仍有挑戰性: – 辨識移動物體 – 使用移動感測器偵測與追蹤移動物體 • 實際應用時,依需求選擇完整求解或簡化求解 – 解答的一致性(consistency) – 計算複雜性(computational complexity) • 與路徑規劃、運動控制器的結合 • 新感測器的發展與應用 • 新的應用領域 Applied Electronics Technology, NTNU 46/48 Visual SLAM of Robot Vacuum Cleaner (Samsung Hauzen RE70V) Applied Electronics Technology, NTNU 47/48 Autonomous Quadrotor Mapping, Localization and Trajectory Following Using LiDAR University of Pennsylvania Applied Electronics Technology, NTNU 48/48