The European Space Agency Tender

The Objective Of This Activity Is To Leverage Machine Learning Methods To Augment The Performance Of Robust Motion Planning And Control Methods Applicable To A Wide Range Of Space Applications, Such As Autonomousplanetary And Asteroid Landing, Spacecraft Rendezvous And Proximity Operations, Among Others.description: The Ability To Autonomously Plan And Execute Highly Constrainedmanoeuvreswill Be A Critical Enabler For Many Future Space Missions Ranging From Planetary Landers, Deep Space Probes, Large Spaceconstellations And Robotic Servicevehicles. By Leveraging Autonomous Planning Capabilities, Spacecraft Can Safely Cope With Dynamic Environments, Failures Scenarios, Complexobjectives, And Stringent Constraints.autonomous Planning And Control Typically Rely Ona Model Of Thespacecraft And Its Environment To Make Predictions About The Future And Optimize Certainobjectives Subject To Various Constraints On The System State And Control Inputs. To Be Useful For Control Or Guidance Purposes, These Computationsneed To Be Performed Within A Short Time And On The Limited Computational Resources Available On Modern Spacecraft. Furthermore, To Guarantee Safetyand Reliability At System Level, Proper Care Should Be Taken Into Account For The Inevitable Physics Modelling And State Estimation Errors.when Prioritizing Factors Other Than Execution Time, One Can Leverage State-of-the-art Optimization-based Heuristics To Seek Viable And High-accuracysolutions For These Intricate Robust Motion Planning Challenges. However, Performing These Computations On-board The Spacecraft In A Reasonableamount Of Time Remains A Challenge For Non-trivial Planning Problems.furthermore, These Heuristics Typically Rely On An Initial Guess Of The Solution That Is Iteratively Refined Using Numerical Methods. Therefore, The Performanceof These Motion Planning Algorithms Can Be Highly Sensitive To The Quality Of The Initial Guess. If A Feasible Initial Guess, (i.e. Satisfying All The Hardconstraints Of The Planning Problem) Cannot Be Readily Generated, Then It Is Impossible To Guarantee That The Algorithm Will Converge To One In A Finiteamount Of Time.recent Advances In Machine Learning Have Made Possible To Alleviate Some Ofthe Limitations Of These Optimization-based Motion Planning Heuristics Anddramatically Enhance Their Performance And Reliability. For Example, Supervised Learning Can Be Used To Compress Vast Amounts Of Optimal Trajectories Obtained Offline Using High Accuracy Solvers Into A Compact Neural Network Representation That Can Be Efficiently Used On-board The Spacecraft.this Neuralnetwork Can Guarantee A Certain Level Of Fitting Error Across Thetraining Dataset And Can Be Used Directly As A Look-up Table Or Be Furtherrefined Through Optimization-based Methods. In This Case, Most Ofcomputational Effort Will Be Spent Off-line During The Data Generation And Neuralnetwork Training Steps. Another Possibility Is To Use Reinforcement Learning Todirectly Tune Parts Of Theguidance And Control Architecture Directly Based Onreward Signals And A Physics Simulator. In This Case, Dispersions In Physical Parameters And Possibly The Uncertainty Propagation Through The Dynamicswould Be Used During The Training Process.in This Context, The Proposed Activity Shall Leverage These Recentadvancements In Machine Learning To Develop A Methodology For Optimalspacecraft Motion Planning And Control That Is Both Computationally Efficient Tobe Deployed On Space Hardware And Can Handle Known Modelling And Stateestimation Uncertainties By Design. The Motion Planning And Control Systemshall Be Appropriately Caged By A Safety Monitor That Ensures Thatthe Controlsignals Used To Drive The Actual Spacecraft Do Not Lead To Constraint Violationseven In The Presence Of These Unavoidable Uncertainties. For Example,reachability Analysis And Covariance Propagation Could Be Used To Ensuresufficientprobabilistic Margins With Respect To The Path Constraints.the Capabilities Of The Novel Motion Planning Technology Shall Bedemonstrated In A Physics Simulator And Compared Against An Establishedheritage Baseline. This Simulator Shall Capture The Key Challenges Of Arelevant, Realistic And Complex Space Mission Scenario Where Alternativeheritage Motion Planning Methods Are Known To Be Insufficient And Where Thenew Methods Can Bring The Most Benefit. Possible Examples Of Such Missionscenarios Include Autonomous Planetary And Asteroid Landing, Spacecraftrendezvous And Proximity Operations, Coupled Attitude And Translation Controlof Underactuatedand/or Low-thrust Spacecraft. The Algorithms Shall Be Stresstestedon The Simulator By Including Uncertainties Both Known And Unknown Attraining-time.finally, The Computational Performance And Reliability Of The Machine-learningbased Motion Planner Shall Be Demonstrated And Benchmarked On An Onboardprocessor Representative Of A Space Missions.the Activity Shall Encompass The Following Tasks: Scenario And Requirements Definition. Development Of A Physics Simulator And A Heritage Baseline Motion Planning Strategy For The Selected Mission Scenario. Development Of Machine Learning-enhanced Robust Motion Planner. Performance Evaluation, Stress Testing On The Simulator. Performance Benchmarking Of The Algorithms On A Representative Space Processor.note: This Activity Is Currently In The Gstp Work Plan But Will Only Be Implemented After Confirmation Of Financial Support From Delegations. The Final List Of Participating States Will Be Known At Tender Issue. Disclaimer : Closing date mentioned is an assumption date, Please contact the tendering authority to get exact closing date.