中文版
English
副研究員  |  劉進興  
 
contact
vita
education
experience
interests
descriptions
activities
honors
grants
publications
software
Personal (New window)
lab (New window)
 
 
 
 
 
Research Descriptions
 

      我們的研究是應用導向, 促成自主系統具有某種層面的自動化。近期的工作集中在(地面或空中)機器人和自駕車相關的實用軌跡計算以符合可行性、安全性和性能要求。路徑規畫與導航是機器人、自駕車、無人機等自主系統的一個最基本的問題之一:搜尋或計算出起點到終點且合乎拘束條件,且優化一個或多個目標函數,能看環境的不確定性和變動而適度調整。自動駕駛(Autonomous driving, AD) 或自動駕駛汽車(AV)作爲未來的交通方式,得經由發展運動規劃的智能化來實現。我們的工作旨在滿足確保AD的軌跡生成的可行性、安全性以及能保證或可預測的性能的需求。重點聚焦在運用優化(數值方法和基於羣體的啓發式搜索)和基於控制的解決策略,解決運動/路徑/軌跡規劃問題;並將機器學習應用到機器人和控制。

我們大多數結果是基於理想的單輪車運動模型(unicycle model)的模擬而得。我們最近進行的自主系統導航技術或概念的可行性驗證有幾個:

        1. 機器人在未知環境中的自主移動與執行保全任務,諸如危險建築中需要量測損毀程度或尋找特定物體、軍方或警方針對可疑或危險區域的偵查、巡邏和攔截入侵者、不宜人類活動的環境中的科學調查與監控、甚至是日漸熱門的居家照護與保全,其最佳解決方案之一就是能即時與環境互動的自走型機器人。

目前我們所使用的自走型機器人寬約30公分,重約15公斤,以左右兩側輪子的兩具伺服馬達驅動,控制其前進後退與原地旋轉,配備雷射掃描器能偵測前方240度範圍內的物體。藉由我們開發的(極座標)導航系統,讓機器人在未知環境中前往給定座標的目的地,並進行指定的任務。而移動的過程中,系統會引導機器人自動繞過障礙物或避免受困於複雜環境。此外,傳統導航系統中,大多會紀錄機器人行走路徑以避免重複打轉,我們採用標記排斥區域的記憶方式達到同樣的功能,卻僅需要極小的記憶體空間。

 

--建造一使用可轉動的雷射測距儀量測水平面與高度距離資料進行室內3D 地圖重建

http://www.youtube.com/watch?v=ceBAgnJhack

--室內導航機器人實驗

目標導向反應式導航的(極座標)路徑循環迴避 http://youtu.be/m19xuV7E9R4,  http://youtu.be/jGuq4SsEN9Q 

--沿牆走及人偵測The system has two primary components, i.e. wall following based on interval type-2 fuzzy logic control and sensing of human in executing trajectories or actions.The human detection sensor allows the effective detection of a close-in human in the wall-following route, thus enabling wall following could be performed in a human friendly manner.

https://www.youtube.com/watch?v=aj_6KkYbbX4

https://www.youtube.com/watch?v=gdGf0CkaoNY

 

2. 三維平滑流形測地線路徑規畫及其在無人飛機飛行路徑規畫的應用

By The ability to plan a path and then to follow the planned path on flat and non-flat terrain in an optimal fashion is central to indoor and outdoor mobile robots navigation. Such a planner requires an accurate geometric 3D environment map that encodes real terrain geometry. We propose a geodesic path planner based on gradient-descent of an energy function that is model independent, analytical and operate on the terrain modeled as a parametric surface or smooth manifold to compute a length minimizing path on a given surface. It generates continuously differentiable locally geodesic paths based on nonlinear geodesic equations determined by surface primitives on which the mobile robot moves. The planner developed is implemented in a planning application and validated through a simulated uav flight that follows the shortest length path towards the goal position.

3.數值調和函數在路徑規劃的應用

Harmonic potential function (HPF) is the solution to two-point boundary value problem of Laplace equation in a closed domain. The streamlines, defined as the gardient of HPF at each point of the domain, serve as the guidance vector field for navigation without trapped in local minima. We develop a HPF-based obsatcle avoidance system that integrates with pure pursuit for transition to another streamline for curvature-constrained mobile robot.To enhance the accuracy of gradient computation, log-space harmonic function is adopted.

4.最短時間軌跡

--本研究聚焦於在特定情境中,車輛沿著生成軌跡,能夠在最短時間之內,從初始狀態到目標狀態。具體而言,在比較各種G2 S形車道切換參數多項式曲線的效率和機操控性,以確定在遵守約束條件的前提下,允許車輛最高的速度進行原點到終點的路徑移動。經過模擬演算的結果顯示,路徑愈短,以及最大曲率愈低的車道變換路徑,車輛(或機器人)移動效率愈高。綜而言之,這個研究成果可望提供車輛駕駛者更安全、更有效的路徑建議方案,特別在緊急救災的情境下,內建路徑軌跡規劃系統的價值將更加凸顯。

--uav(autonomous flight technology):The broad dissemination of unmanned aerial vehicles (UAV s), specifically quadrotor aircraft, has accelerated their successful use in a wide range of industrial, military, and agricultural applications.Dubins vehicle is an ideal kinematic model of constant speed, constant altitude uav.The trajectories for uav must be safe and flyable. Dubins車在有或無障礙物環境的最短時間軌跡,我們採取Hamiton-Jacobi方程式的fast sweeping method數值方法求解,並與解析解比較其精確度和收斂性。This part of work broadly aligns with the numerical optimal control framework that implements trajectory optimization.

--增強學習(Reinforcement Learning)在理論、演算法及應用取得長足的發展與進步。然而增強學習在機器人與自動控制領域的應用仍有待研究開發,其阻礙有二:一、機器人與自動控制通常為連續空間及時間的高維度任務,二、此種任務經常帶有非線性的狀態與控制限制,因此增加了相關應用的困難度和挑戰性。

We consider a data-driven problem of inferring the near time-optimality of both the control and trajectory from observations of their rewards in executing a path following task.
Aim: Learn the (feedback) control function for the vehicle with bounds on longitudinal acceleration to follow a path to reach a goal as fast as possible despite the uncertainty (inaccuracies) in model parameters.

本研究將一套既有的演算法(PILCO)應用於時間優化速度控制問題中,取得階段性的成功。我們用高斯過程(Gaussian Processes)歸納阻尼式雙積分器(damped double integrators) 機器人行為模型,預測未來路徑並基於此預測做出優化。此套演算法被成功地模擬於電腦中和測試於實體的自動小車上,以極少量的數據達成最佳時間控制。未來相關研究可著重應用此方法於更高維度的任務,並且引入模型預測控制(Model Predictive Control)以因應非線性限制。

Code is available at https://github.com/brianhcliao/PILCO_AlphaBot

 

5.TSPN and TSP-CPP

覆蓋路徑規劃(CPP)被廣泛應用在製造、清潔、掃雷、割草,以及無人機地圖測繪、氣候與氣象監測、SAR以及空氣品質監測等各項實務工作上。對於已知有障礙物的環境,一種有效的解決方案,是將環境分解成若干個單元,將擬測區域被所有組成單元覆蓋。然後,透過單元的訪問順序的確定,將所有單元內部的路徑連接在一起。最後找到途經每個單元,並返回原始單元的最短跨單元路徑,完成避障物的路徑探索,這個解決方案的構想,類似於旅行商問題;而額外的變化,是每個單元都有多個單元內部路徑,因為每個單元的進入和退出點不同,從而影響跨單元路徑的產生。

這個聚合型的旅行推銷員,和覆蓋路徑規劃問題,被稱為TSP-CPP,類似於有鄰居的TSPTSPN)。它具有許多實際應用,例如清潔機器人、SAR和通過四旋翼無人機進行空氣品質監測。例如,SAR搜索和監測需要覆蓋大面積的快速計算。在優化TSPN路徑的距離時,航點的數量和位置以及其訪問順序均起著重要作用。先前的研究聚焦于開發一種基於GATSPN求解器(https://github.com/shao-you/CBGA),它可以適應半徑和重疊數量不同的鄰域。解決TSP-CPP,需要同時確定訪問節點的訪問順序,以及每個訪問節點的轉換點。近期解決TSP-CPP的方法包括為TSP調整的動態規劃法(DP),或者使用窮舉搜索每個單元的進出點的組合,然後使用TSP求解器解決每個進出點的組合。這兩種方法都具有指數複雜度,對於具有大量單元的複雜環境是成效是受到侷限的。GA 可以使用來解決各種TSPN變異。我們提出了一種基於GATSP-CPP解決方案,旨在克服DP所帶來的維度限制,以實現時間效率和路徑最佳性之間的良好平衡。這項工作在TSPN基礎上展開,並引入新的初始種群生成方法以及更高效的交配和突變算子的組合。我們的方法被證明在所有模擬環境中都可以找到真正的最佳解決方案,對於被分解成大量單元的地圖要快100倍。 此外,它可以在有限的資源條件下獲得最佳解和中間解。計畫在未來將該方法擴展到Dubins'' TSPNDTSPN)。

GA codes for TSP-CPP  https://github.com/WJTung/GA-TSPCPP

6.collision detection

Collision detection in robotics has arisen primarily in the context of obstacle avoidance and path planning, in which robot geometry is tested for collision with geometric models of environment obstacles along the trajectory followed by the robot. For the application of assembly or path planning, the most expensive part of planning is ensuring that there is no collision between objects/ parts.

-Exact collision detection for scaling tolerance

In this work, we deal with collision detection of a pair of convex polyhedral with the variability in size (dimension) but not shape, i.e. scaled convex polyhedra.A methodology for quickly verifying two convex polyhedral parts, each with its size scaled by a factor, could be mated by identifying conditions under which we can achieve desirable mating configurations for toleranced polyhedral parts.

-Swept volume reconstruction using deep learning model of signed distance function

產業自動化愈發普及,機械手臂也愈加常見。為了優化這項科技,人類與機器人協作(HRC)這種結合人類決策能力與機器人可程式化的方式得以發展,以達到更高的生產力和靈活度。透過使用HRC系統,人類員工與機器人可以在共享的工作空間內協力完成共同的目標。這種作法能夠充分發揮人類員工和機器人各自的優勢,互相補足彼此的能力。當機器人進入與人類夥伴互動的工作場所時,安全是至關重要的考慮因素,必須採取措施減少碰撞風險,確保工人的安全。為了實現這一目標,精確的碰撞偵測和採用基於控制或靈活運動規劃的有效避障策略是必不可少的。優先考慮工人的安全對成功的人機協作至關重要,這能夠讓雇主在不危及員工健康的前提下有效地利用自動化技術。

 

HRC system (e.g. robot-assisted surgery, feeding or dressing, etc. ) that makes the human and the robot working together in a shared workspace to complete a common task has been shown to increase considerably the overall capabilities of the system.For robotic applications involving interaction with human as partners such as HRC system, the ability to detect unsafe situations is crucial for HRC system, sv (or its bounding volumes ) enables continuous collision prediction at each time step at each pose of the arm along the entire or partial (maybe human following with awareness of human motion intention and prediction ) trajectory, so that when the robot arm reaches certain portion of the workspace is known and thus can trigger an emergency stop to react to unexpected events for safety guarantee.Human-robot collaborations can be assured to be safe by means of sv with significant reduction of number of collision calls. 

Given an articulated robot arm ( a kinematic chain) S, we are given the boundary surface of the volume which is swept by S following a trajectory.The surface is defined by  {f : R3 → R , f = 0}  as its zero- level-set. We want to find the implicit neural representationswhich encode a surface as the level set of a multilayer perceptrons neural network mapping 3D spatial coordinates to implicit function valuesof the boundary of swept volume. Given a function f whose zero level-set defines a geometric surface S, we want to compute the signed distance to S (or an approximation), while preserving exactly the zero level-set of f, since an SDF measures distance to the nearest surface point in any direction,  which implicitly represents object shape. Then, reconstruct the swept volume by the learned sdf to validate the accuracy in terms of Hausdorff distance to the actual sv and verify feasibility of using sdf for representing the given swept volume implicitly. We examine the accuracy of such a sv reconstruction approach using deep learning with a variety of robots tracking complex trajectories in 3D shared workspaces.The results show consistency between the learned sdf and sv with remarkably accurate reconstruction, That is, it is feasible to build a consistent, SDF-based (distance-preservation) , memory efficient reconstruction of sv via an implicit neural representation. 

Since the safety of the path must be guaranteed before it is actually executed  by the robot or during execution, we propose leaned-SDF -guided  continuous collision detection for time-efficient task execution, allowing the manipulator to foresee over the trajectory duration and avoid possible collisions by altering its planned motion..We can identify all possible collisions between each of the links and human, therefore manipulability area or  the dangerous operational areas, based on learned sdf as an implicit representation of sv along a given trajectory 

demo of learned-sdf-guided continuous collision detection for pick-and-place task: https://youtu.be/dDvTaYYy9mk

An implementation of DRRT with SDF on GPU achieves 4x -62x faster path computation in a simulation in a simple dynamic environment with one moving AGV ,depending on AGV speed.

https://youtube.com/shorts/W1dtQ3viSEg?feature=share

 

 

7. G3 smooth path design

G3 curves ensure comfort and smooth change in curvature derivative and acceleration. G3 Path/motion/ Trajectory planning is concerned with finding or computing or designing a subclass of all trajectories (of the the unicycle ) that connect the given terminal states/poses/configurations (I.e. meet the interpolation constraints) by composing path/motion primitives at some time instants (traversed time) with third  order geometric continuity, where the orientation, curvature and curvature derivative at the starting and end points are prespecified.

To account for arbitrary nonzero curvature and arbitrary curvature derivative at either one or both ends of the curve and incorporation of constraints for handling more complex driving scenarios and road layouts, we proceed by the use of a single directional (forward), smooth turning path based on 7th degree Bézier curves to meet the curvature constraint. The family of 7th degree Bézier curves in our design is equivalent to an eta-3 spline interpolating imposed boundary conditions with control points defined by two parameters of end speeds.An advantage of using the 7th degree Bézier form is we can generate various paths through selection of suitable parameter values  in a more intuitive manner. Specifying  nonzero boundary curvature and  incorporating collision avoidance at the  same time can be achieved more easily and intuitively to some extent via the Bézier form , and its closed-form provides a low  complexity of trajectory computation. It is used for designing the maneuvers of entering the roundabout or in-roundabout lane change by placing control points taking into account the entire lane width of  a circulatory roadway. In this computationally simpler approach, the starting velocity and the ending velocity of the forward (single-directional) trajectory are the design parameters. The feasible space of path-defining parameters  is reduced to a box. This allows us to find a feasible trajectory  by iteratively searching the feasible parameter space for determining/ refining or direct optimization the control points of its natural equivalent 7th degree Bézier curve for a given  maneuver time (the duration to execute the trajectory) T until a satisfactory G3 trajectory is found.

 


 
 
bg