COLREGS Based Collision Avoidance of Unmanned Surface Vehicles

Transcription

1 COLREGS Based Collision Avoidance of Unmanned Surface Vehicles Aug 2nd, 2017 Michael R. Benjamin Kyle Woerner, Michael Novitzky, John Leonard MIT Dept. of Mechanical Engineering Computer Science and Artificial Intelligence Lab Acknowledgements Office of Naval Research (ONR 333) has sponsors the COLREGS algorithm development. Dr. Robert Brizzolara. Office of Naval Research (ONR 311) has sponsored MOOS-IvP for nearly 10 years. Dr. Don Wagner, Dr. Behzad Kamgar-Parsi, Dr. Wen Masters. Battelle has sponsored, beginning in 2010, MOOS-IvP, MOOS, the development of our autonomy course at MIT, the purchase of the Heron platforms, and the build-out of our Marine Autonomy Bay facilities. Mr. Mike Mellott and Mr. Mike Kuhlman. Thales UK office has sponsored the COLREGS work, beginning in Faculty collaborators: Prof. Henrik Schmidt MIT MechE Prof. John Leonard MIT MechE/CSAIL Prof. Paul Newman Oxford 1

2 COLREGS Autonomy Idea: Enable autonomous surface vehicles to obey the Rules of the Road COLREGS. Establish a road test for validating the autonomous collision avoidance. COLREGS Autonomy Funded by: Office of Naval Research (ONR) Idea: Enable autonomous surface vehicles to obey the Rules of the Road COLREGS. Establish a road test for validating the autonomous collision avoidance. Research Focus: Map the protocols written for humans into algorithmic format. Find minimal set of field tests that validate widest set of scenario permutations. Crossing Situation Head-on Situation Technical Approach: Collision avoidance protocols mapped to set of modes, and sub-modes. Modes map to a unique form of objective function. Multi-objective optimization with IvP to solve. Impact: Autonomous long-duration coastal sampling with autonomous platforms. Multi-vehicle/swarm capabilities can be built on COLREGS foundation. 2

3 Payload UUV Autonomy (3 Architecture Principles) Payload Computer Main Vehicle Computer Payload Computer IvP Helm MOOS Middleware MOOS Applications Behavior-Based Modular HELM Autonomy Research Focus Areas Aquaticus RoboChallenge COLREGS Training/Documentation/V&V MOOS-IvP 3

4 Overview of the IvP Helm Behavior Output and Action Selection 1 Mail is read in the MOOS OnNewMail() function and applied to a local buffer. 2 The helm mode is determined, and set of running behaviors determined. 3 Behaviors execute posting MOOS variables and an IvP function. 4 Competing behaviors are resolved with the IvP solver. 5 The Helm decision and any behavior postings are published to the MOOSDB. Overview of the IvP Helm Behavior Output and Action Selection Behavior 1 Behavior 2 Behavior 3 f 1 (x 1, x 2,, x n ) f 2 (x 1, x 2,, x n ) f 3 (x 1, x 2,, x n )! x * =! argmax x k 1 i= 0 w i f i (! x ) Action Objective Functions Solver The objective functions are called IvP functions functions of a certain format. The Solver is called the IvP Solver they exploit the IvP function structure. Typical Decision Space: Heading, Speed, Depth 4

5 IvP Functions The IvP Function vs. Underlying Function An IvP function is a piecewise linear approximation of an objective function, over a discrete decision space (domain). Piecewise Linear Approximation 525 Pieces Underlying Function Interval Programming Solution Algorithms Overview An IvP problem consists of a set of k functions, each with a priority weighting. The solution is given by:! x * =! argmax x k 1 i= 0 w i f i (! x ) f 1 f 2 f 3 The Search Tree: 1 level for each function n k leaf nodes (n pieces per function). The Solution algorithm: Branch and bound Pruning based on intersection lookahead. 5

6 Overview of the IvP Helm Behavior Output and Action Selection Behavior 1 Behavior 2 Behavior 3 f 1 (x 1, x 2,, x n ) f 2 (x 1, x 2,, x n ) f 3 (x 1, x 2,, x n )! x * =! argmax x k 1 i= 0 w i f i (! x ) Action Objective Functions Solver Example: Waypoint Traversal Behavior Overview of the IvP Helm Behavior Output and Action Selection Behavior 1 Behavior 2 Behavior 3 f 1 (x 1, x 2,, x n ) f 2 (x 1, x 2,, x n ) f 3 (x 1, x 2,, x n )! x * =! argmax x k 1 i= 0 w i f i (! x ) Action Objective Functions Solver Example: Collision Avoidance Behavior 6

7 Overview of the IvP Helm A Simple Collision Avoidance Example Behavior 1 Behavior 2 Behavior 3 f 1 (x 1, x 2,, x n ) f 2 (x 1, x 2,, x n ) f 3 (x 1, x 2,, x n )! x * =! argmax x k 1 i= 0 w i f i (! x ) Action Objective Functions Solver COLREGS Collision Avoidance (sponsored by ONR, Thales UK) The objective is to avoid collisions with other autonomous and non-autonomous vehicles. COLREGS are the rules of the road for seagoing vessels. They provide a protocol of roles and required actions between vessels. They were written for humans, not autonomous systems. Collision Avoidance WITHOUT COLREGS 7

8 What s Wrong with Non Protocol Based Collision Avoidance Consider the head-on situation If ownship rates candidate maneuvers based on closest point of approach, a maneuver to port or starboard looks equally good. 0 Ownship m/s 180 What s Wrong with Non Protocol Based Collision Avoidance Consider the head-on situation If ownship rates candidate maneuvers based on closest point of approach, a maneuver to port or starboard looks equally good. And the same is true for the contact, so It s possible one turns to port and the other to starboard, or vice versa. Contact 0 0 Ownship m/s m/s

9 COLREGS Collision Avoidance The head-on situation is referenced in Rule 14 of the COLREGS. When two power-driven vessels are meeting on a reciprocal or nearly reciprocal courses so as to involve a risk of collision each shall alter her course to the starboard so that each shall pass on the port side of the other The COLREGS IvP Behavior on ownship heavily penalizes the wrong kind of turn. 0 Ownship m/s 180 COLREGS Collision Avoidance The head-on situation is referenced in Rule 14 of the COLREGS. When two power-driven vessels are meeting on a reciprocal or nearly reciprocal courses so as to involve a risk of collision each shall alter her course to the starboard so that each shall pass on the port side of the other The COLREGS IvP Behavior on ownship heavily penalizes the wrong kind of turn. And, of course, so does the contact: Contact 0 0 Ownship m/s m/s

10 COLREGS Collision Avoidance The give-way (crossing) situation is referenced in Rule 15 of the COLREGS. When two power-driven vessels are crossing so as to involve risk of collision, the vessel which has the other on her own starboard side shall keep out of the way and shall, if the circumstances of the case admit, avoid crossing ahead of the other vessel. The give-way vehicle may cross ahead of the other vessel if it clearly makes more sense. 0 Ownship 2.4 m/s m/s 4 m/s 180 COLREGS Collision Avoidance (sponsored by ONR) The objective is to avoid collisions with other autonomous and non-autonomous vehicles. COLREGS are the rules of the road for seagoing vessels. They provide a protocol of roles and required actions between vessels. Collision Avoidance WITH COLREGS 10

11 Testing / Validation Simulation In-Water Tests Testing / Validation 11

12 Testing / Validation Testing / Validation 12

13 How do we learn from the data? The Challenge: 12 hours of simulation 4 vehicles 1009 encounters RESULTS 1009 encounters 2 collisions -one 0.95m -one 7.5m 9 near misses (8-12m range) Auto Generated Encounter Plots collision Encounter Plots Part of the alogview tool Open Source Works on any mission log file Developed Nov 15 to present. In next release of MOOS-IvP collision Near miss range clear 13

14 In-Water Tests Heron M300 (4) Human-operated Mokai 27 In-Water Tests Day mode 28 14

15 In-Water Tests Day mode 29 In-Water Tests 30 15

16 COLREGS Collision Avoidance (sponsored by ONR, Thales UK) Open Question: Does COLREGS Scale to larger sets of vehicles, swarms? END 16