AMRs, AGVs, and Machine Learning for Multi-Robot Coordination

Invited Talk given at Zukunftskongress Logistik – 42, AI24 Llamar Institute.

Transcript

1. Jean-Marc Alkazzi Engineering Manager - Applied AI @ idealworks AMRs, AGVs, and Machine Learning for Multi-Robot Coordination AI24 – Lamarr Institute September 5th, 2024 Session 5: AI and Simulation Technologies in Intralogistics (Workshop Room 5)
2. AI and Simulation Technologies in Intralogistics • AI and Simulation are tools to help solve specific challenges. • To understand how to use the right tool for the job, we need to clearly define the job. • The job in question is related to Intralogistics. • Problem Definition: Given a number of robots tasked to autonomously transport goods inside a logistics facility, we need to coordinate their behavior by • Assigning the right task to the right robot, to accomplish them at the correct time, in the correct sequence while minimizing total travel time. • Assigning a route per robot so that each robot safely reaches its destination in the fastest time possible while avoiding potential congestions. • Ensure the maximum availability of robots while warranting an extended battery lifetime. 2
3. Industrial Environments: Expectations 3 https://www.belvoir.net/wp- content/uploads/2022/10/iStock-1138429558-scaled.jpeg https://www.wealthmanagement.com/sites/wealthmanagement.com/files/st yles/article_featured_retina/public/co-warehousing.jpg?itok=xEH8z4Yk https://alta-commercial.com/wp- content/uploads/2020/06/Warehouse-Building.png
4. Industrial Environments: Reality 4 Inside a BMW Group Production Environment
5. AI and Simulation Technologies in Intralogistics • Intelligence • Single-Agent vs Multi-Agent intelligence • Decentralized vs Centralized • Responsibility (agent vs central system) based on Decision • Simulation • A tool for understanding and experimenting • Simulation fidelity requirements • What can/should be reliably simulated? 5
6. Jean-Marc Alkazzi Engineering Manager - Applied AI @ idealworks AMRs, AGVs, and Machine Learning for Multi-Robot Coordination and how AI and Simulation can help AI24 – Lamarr Institute September 5th, 2024 Session 5: AI and Simulation Technologies in Intralogistics (Workshop Room 5) Challenges of
7. Key topics for this talk • Decentralized vs. Centralized Intelligence: Finding the right Balance in Robotic Systems • Leveraging Machine Learning for Multi-Agent Path Finding • The Role of Simulation in Developing Robust Autonomous Ecosystems 7
8. Key topics for this talk • Decentralized vs. Centralized Intelligence: Finding the right Balance in Robotic Systems • Leveraging Machine Learning for Multi-Agent Path Finding • The Role of Simulation in Developing Robust Autonomous Ecosystems 8
9. Industrial Robot Fleet Management System • To tackle the defined problem, an orchestrating system is used. At idealworks, it is AnyFleet. • The Fleet Management System (FMS) is responsible for deciding • Which robot is assigned to which task • Which route should each robot take to achieve its task • When should each robot go to a charging/parking spot to ensure 24/7 operation • … 9
10. Industrial Robot Fleet Management System • Where to shift the intelligence based on the decision to be made? • When should the robot decide on its own, when should it ask for permission, and how to decide on the required responsibilities? 10
11. idealworks' Approach idealworks' Approach 11 Reaction Speed Observation Space Local Global Slow Fast
12. idealworks' Approach idealworks' Approach 12 Reaction Speed Observation Space Local Global Slow Fast
13. idealworks' Approach idealworks' Approach 13 Reaction Speed Observation Space Local Global Slow Fast
14. idealworks' Approach idealworks' Approach 14 Reaction Speed Observation Space Local Global Slow Fast
15. idealworks' Approach idealworks' Approach 15 Reaction Speed Observation Space Local Global Slow Fast
16. idealworks' Approach idealworks' Approach 16
17. idealworks' Approach idealworks' Approach 17
18. Why is a Fleet Manager even needed? Can't we just have smarter robots and call it a day? The Existential Question
19. Aren't Smarter Robots enough? 19 It works, but definitely without any safety guarantees.
20. Aren't Smarter Robots enough? 20 [2024] Waymo Self-Driving Cars Humans (supposedly the smartest species) More freedom ≠ More efficiency
21. If smarter robots aren't enough, how smart should a robot be?
22. If smarter robots aren't enough, how smart should a robot be? It depends...
23. The AMR vs AGV war • AMR = Autonomous Mobile Robot • AGV = Automated Guided Vehicle • Side 1: Always use smart AMRs! AMRs are the future! • Side 2: You don't need intelligence with a good centralized system! Use AGVs for deterministic behavior! 23
24. Use the right robot for each job and take decisions at the correct layer.
25. What makes it more challenging in Industrial Environments? • Integration & Control • Limited control over integrated robots’ autonomy • Diverse decision-making processes across FMS based on providers • Lack of widespread robot-to-robot collaboration standards (Layer 2) • Latency Issues • Observation gathering delays • Reaction time delays • Execution delays in chaotic environments 25
26. What makes it more challenging in Industrial Environments? • Operational Complexities • Heterogeneous fleets with varying "intelligence requirements" • Orchestrating a fleet in the presence of uncontrollable vehicles (old AGVs, human-driven vehicles) • 24/7 operational challenges • External Factors • Poor internet connectivity • Other vehicles colliding into your fleet • Unexpected environment layout changes 26
27. Key topics for this talk • Decentralized vs. Centralized Intelligence: Finding the right Balance in Robotic Systems • Leveraging Machine Learning for Multi-Agent Path Finding • The Role of Simulation in Developing Robust Autonomous Ecosystems 27
29. Leveraging Machine Learning for Multi-Agent Path Finding 29 Figure A: The League of Robot Runner Competition. Sponsored by Amazon Robotics.
30. Leveraging Machine Learning for Multi-Agent Path Finding 30
31. Leveraging Machine Learning for Multi-Agent Path Finding • Representation • Representation for Planning • Environment Optimization • Representation for Selection • Planning • Augmenting Existing Solvers • Developing Decentralized Policies • Execution 31
32. ML for MAPF: Representation for Planning 32 Figure B: Cooperative Timed Roadmaps (CTRMs) generation. Figure taken from (Okumura et al., 2022) Figure A: Representation Trade-Offs
33. ML for MAPF: Representation for Environment Optimization 33
34. ML for MAPF: Representation for Selection 34 Figure A: General Approach for MAPF Algorithm Selection Figure B: Ren et al. 2021 Input Encoding
35. ML for MAPF: Augmenting Existing Planners 35 Figure A: Learning an agent prioritization policy.
36. ML for MAPF: Developing Decentralized Policies 36 (With communication)
37. ML for MAPF: Execution Figure A: Congestion Delays Prediction architecture. Figure taken from Yu et al. 2023 37
38. Key topics for this talk • Decentralized vs. Centralized Intelligence: Finding the right Balance in Robotic Systems • Leveraging Machine Learning for Multi-Agent Path Finding • The Role of Simulation in Developing Robust Autonomous Ecosystems 38
39. The Role of Simulation in Developing Robust Autonomous Ecosystems • Testing any hypothesis we previously discussed is time/resource consuming. • To stay ahead, we need to increase our development iteration speed. • Grid simulation helps you test faster, but does not reflect the real world (unless you have a structured field e.g., Amazon Robotics with floor QR codes) • We need to simulate what makes our problem challenging • Delays (Observation, Reaction, Execution) • Chaos • Sudden failures of robots or FMS services • … 39
40. The Role of Simulation in Developing Robust Autonomous Ecosystems Layered approach to using Simulation • Statistically modelling the environment • Simple but scalable 2D Simulation (Focus on testing AnyFleet Services) • Complex but realistic 3D Simulation (Focus on testing robots and final e2e) ALL of them are important! 40
41. The Role of Simulation in Developing Robust Autonomous Ecosystems Figure A: Simulation trade-offs. 41
42. Scalable 2D Simulation 42
43. Realistic Simulation 43
44. Where can you find more? [Stay for the Panel Discussion!] 44
45. How to stay in touch? Let's collaborate! jean-marc.alkazzi@idealworks.com I have fun writing sometimes at https://blog.jeremarc.com where you can also find links to other talks. Not a social media fan, but I use Twitter @jeanmarcalkazzi 45