ai & robotics: [1ex] researchintroduction

16
AI & Robotics: Research Introduction Marc Toussaint Technical University of Berlin Summer 2020

Upload: others

Post on 05-Feb-2022

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: AI & Robotics: [1ex] ResearchIntroduction

AI amp RoboticsResearch

Introduction

Marc ToussaintTechnical University of Berlin

Summer 2020

Outline

bull Brief self-introduction amp research

bull Course Information

bull Go through conferences and papers

2

Short Bio

bull Stationsndash Since 2012 Prof U Stuttgart introductory teaching (AI ML Robotics)ndash 201718 1yr MIT sabbatical 3 months manager at Amazonndash PhD in Evolutionary Algorithms() and Neural Networks()ndash After that Doing rdquoproper thingsrdquo PostDoc in ML in Edinburghndash Emmy Noether group in Berlin (TU Berlin amp FU Berlin)

bull Researchndash Naively Learning vs Thinking (generalization representations

computation)ndash Probabilistic Inference Decision Theory Markov Decision Processes

Reinforcement Learning POMDPs factored and relationalrepresentations in all these

ndash Later Consider much more the REAL worldndash representations amp abstractions for the real worldndash ROBOTICS control motion MANIPULATIONndash General Purpose Physical Reasoning

Reason about anything doable in a Newtonian world

3

Short Bio

bull Stationsndash Since 2012 Prof U Stuttgart introductory teaching (AI ML Robotics)ndash 201718 1yr MIT sabbatical 3 months manager at Amazonndash PhD in Evolutionary Algorithms() and Neural Networks()ndash After that Doing rdquoproper thingsrdquo PostDoc in ML in Edinburghndash Emmy Noether group in Berlin (TU Berlin amp FU Berlin)

bull Researchndash Naively Learning vs Thinking (generalization representations

computation)ndash Probabilistic Inference Decision Theory Markov Decision Processes

Reinforcement Learning POMDPs factored and relationalrepresentations in all these

ndash Later Consider much more the REAL worldndash representations amp abstractions for the real worldndash ROBOTICS control motion MANIPULATIONndash General Purpose Physical Reasoning

Reason about anything doable in a Newtonian world3

Physical Reasoning amp Manipulation

Battaglia Hamrick amp Tenenbaum PNASrsquo13

(Wolfgang Kohler 1917)

bull What are computational models for physical reasoning

bull Reason about anything doable in a Newtonian world4

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 2: AI & Robotics: [1ex] ResearchIntroduction

Outline

bull Brief self-introduction amp research

bull Course Information

bull Go through conferences and papers

2

Short Bio

bull Stationsndash Since 2012 Prof U Stuttgart introductory teaching (AI ML Robotics)ndash 201718 1yr MIT sabbatical 3 months manager at Amazonndash PhD in Evolutionary Algorithms() and Neural Networks()ndash After that Doing rdquoproper thingsrdquo PostDoc in ML in Edinburghndash Emmy Noether group in Berlin (TU Berlin amp FU Berlin)

bull Researchndash Naively Learning vs Thinking (generalization representations

computation)ndash Probabilistic Inference Decision Theory Markov Decision Processes

Reinforcement Learning POMDPs factored and relationalrepresentations in all these

ndash Later Consider much more the REAL worldndash representations amp abstractions for the real worldndash ROBOTICS control motion MANIPULATIONndash General Purpose Physical Reasoning

Reason about anything doable in a Newtonian world

3

Short Bio

bull Stationsndash Since 2012 Prof U Stuttgart introductory teaching (AI ML Robotics)ndash 201718 1yr MIT sabbatical 3 months manager at Amazonndash PhD in Evolutionary Algorithms() and Neural Networks()ndash After that Doing rdquoproper thingsrdquo PostDoc in ML in Edinburghndash Emmy Noether group in Berlin (TU Berlin amp FU Berlin)

bull Researchndash Naively Learning vs Thinking (generalization representations

computation)ndash Probabilistic Inference Decision Theory Markov Decision Processes

Reinforcement Learning POMDPs factored and relationalrepresentations in all these

ndash Later Consider much more the REAL worldndash representations amp abstractions for the real worldndash ROBOTICS control motion MANIPULATIONndash General Purpose Physical Reasoning

Reason about anything doable in a Newtonian world3

Physical Reasoning amp Manipulation

Battaglia Hamrick amp Tenenbaum PNASrsquo13

(Wolfgang Kohler 1917)

bull What are computational models for physical reasoning

bull Reason about anything doable in a Newtonian world4

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 3: AI & Robotics: [1ex] ResearchIntroduction

Short Bio

bull Stationsndash Since 2012 Prof U Stuttgart introductory teaching (AI ML Robotics)ndash 201718 1yr MIT sabbatical 3 months manager at Amazonndash PhD in Evolutionary Algorithms() and Neural Networks()ndash After that Doing rdquoproper thingsrdquo PostDoc in ML in Edinburghndash Emmy Noether group in Berlin (TU Berlin amp FU Berlin)

bull Researchndash Naively Learning vs Thinking (generalization representations

computation)ndash Probabilistic Inference Decision Theory Markov Decision Processes

Reinforcement Learning POMDPs factored and relationalrepresentations in all these

ndash Later Consider much more the REAL worldndash representations amp abstractions for the real worldndash ROBOTICS control motion MANIPULATIONndash General Purpose Physical Reasoning

Reason about anything doable in a Newtonian world

3

Short Bio

bull Stationsndash Since 2012 Prof U Stuttgart introductory teaching (AI ML Robotics)ndash 201718 1yr MIT sabbatical 3 months manager at Amazonndash PhD in Evolutionary Algorithms() and Neural Networks()ndash After that Doing rdquoproper thingsrdquo PostDoc in ML in Edinburghndash Emmy Noether group in Berlin (TU Berlin amp FU Berlin)

bull Researchndash Naively Learning vs Thinking (generalization representations

computation)ndash Probabilistic Inference Decision Theory Markov Decision Processes

Reinforcement Learning POMDPs factored and relationalrepresentations in all these

ndash Later Consider much more the REAL worldndash representations amp abstractions for the real worldndash ROBOTICS control motion MANIPULATIONndash General Purpose Physical Reasoning

Reason about anything doable in a Newtonian world3

Physical Reasoning amp Manipulation

Battaglia Hamrick amp Tenenbaum PNASrsquo13

(Wolfgang Kohler 1917)

bull What are computational models for physical reasoning

bull Reason about anything doable in a Newtonian world4

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 4: AI & Robotics: [1ex] ResearchIntroduction

Short Bio

bull Stationsndash Since 2012 Prof U Stuttgart introductory teaching (AI ML Robotics)ndash 201718 1yr MIT sabbatical 3 months manager at Amazonndash PhD in Evolutionary Algorithms() and Neural Networks()ndash After that Doing rdquoproper thingsrdquo PostDoc in ML in Edinburghndash Emmy Noether group in Berlin (TU Berlin amp FU Berlin)

bull Researchndash Naively Learning vs Thinking (generalization representations

computation)ndash Probabilistic Inference Decision Theory Markov Decision Processes

Reinforcement Learning POMDPs factored and relationalrepresentations in all these

ndash Later Consider much more the REAL worldndash representations amp abstractions for the real worldndash ROBOTICS control motion MANIPULATIONndash General Purpose Physical Reasoning

Reason about anything doable in a Newtonian world3

Physical Reasoning amp Manipulation

Battaglia Hamrick amp Tenenbaum PNASrsquo13

(Wolfgang Kohler 1917)

bull What are computational models for physical reasoning

bull Reason about anything doable in a Newtonian world4

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 5: AI & Robotics: [1ex] ResearchIntroduction

Physical Reasoning amp Manipulation

Battaglia Hamrick amp Tenenbaum PNASrsquo13

(Wolfgang Kohler 1917)

bull What are computational models for physical reasoning

bull Reason about anything doable in a Newtonian world4

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 6: AI & Robotics: [1ex] ResearchIntroduction

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 7: AI & Robotics: [1ex] ResearchIntroduction

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 8: AI & Robotics: [1ex] ResearchIntroduction

Why is this interesting to study

bull Physical Reasoning is under-researchedndash Lots of methodologies for physical modelling but not reasoningndash Focus of main-stream RL specific skills rarr generalization to anything

conceivable in a Newtonian worldndash Robotics task and motion planningndash Cognitive Science needs models

bull Core challenge in robotics

5

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 9: AI & Robotics: [1ex] ResearchIntroduction

Inverting Physics

bull In analogy to inverting graphicsGiven desired outcomes what inputs do we have to send to physics

bull Differentiable Physicsndash Todorov A convex smooth and invertible contact model for trajectory

optimization ICRArsquo11ndash de Avila Belbute-Peres amp Kolter A Modular Differentiable [] Physics

Engine NIPSrsquo17 workshopndash Mordatch et al Discovery of complex behaviors through contact-invariant

optimization TOGrsquo12ndash Note Local() differentiation through KKT conditions of constrained

optimization

bull Gradients are powerful but can they alone solve our problemndash would contradict known complexity of task and motion planningndash rsquozero gradientsrsquo or local optimandash discrete decisions translate to combinatorics of local optima

6

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 10: AI & Robotics: [1ex] ResearchIntroduction

vids

bull Describing Physics For Physical Reasoning Force-based SequentialManipulation Planninghttpswwwyoutubecomwatchv=YxKuVit_23E

bull Differentiable Physics and Stable Modes for Tool-Use and ManipulationPlanning httpswwwyoutubecomwatchv=-L4tCIGXKBE

bull MorehttpswwwyoutubecomchannelUC9ANVqaEC0iM9aZr88sHK3A

these show what we made progress with in the last years

7

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 11: AI & Robotics: [1ex] ResearchIntroduction

Planning rarr Execution

bull So far LGP only describes how to compute plans ndash execution of theseplans is a different beast

rarr Learning to Execute Plansndash Learning to do as planned to revise plans and modelsndash ldquoTrial-and-Error Learningrdquo (literally not just RL)

8

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 12: AI & Robotics: [1ex] ResearchIntroduction

IntCDC

bull Excellence Cluster in Integrated Computational Design andConstruction

9

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 13: AI & Robotics: [1ex] ResearchIntroduction

Questions

10

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 14: AI & Robotics: [1ex] ResearchIntroduction

Official Module Descriptionbull Learning Outcomes

ndash The students understand core research questions and methodologicalapproaches that state-of-the-art publications in AI and robotics currentlyaddress and follow

ndash They can identify the limitations of state-of-the-art researchndash They can constructively discuss what novel methodological research

might lead to advances in fundamental open research questionsndash They understand how modern scientific publications in the field are

structured and how literature search is performed efficiently

bull Contentndash The subject matter are scientific papers either classical seminal papers

or papers from current AI and robotics conferencesndash The papers are selected to represent core approaches towards real-world

AI for instance model-free and model-based RL classical planning andcontrol real-world- and robotics-centric approaches sim2real andreal2sim exploiting abstractions and hierarchies

ndash The exercises require students to read scientific papers watch relatedvideos discuss and find relatedcompeting literature (re-)formalize theexact problem formulation of papers and potentially reproduce results

ndash The lectures (this is not a seminar) introduce further background anddiscuss the papers in a larger scientific context

bull Maximum capacity 25

11

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 15: AI & Robotics: [1ex] ResearchIntroduction

Official Module Description

bull Desirable prerequisites for participation in the courses Studentsshould have

ndash This course is intended for senior MSc students interested in becoming aresearcher in AI or robotics

ndash In depth knowledge in robotics (passed a robotics course)ndash General knoweldge in AI and machine learning

bull 180h (6LP)

bull Portfolio examinationndash Essay on a lecture topic 50 writtenndash Written test 50 written 60 minutes

12

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13

Page 16: AI & Robotics: [1ex] ResearchIntroduction

Organization

bull See https

wwwusertu-berlindemtoussainew-at-tu-berlinhtml

bull Teaching Assistantsndash Jung-Su Ha (PostDoc)ndash Ingmar Schubert (1yr PhD)

bull Course Webpagehttpswwwusertu-berlindemtoussaiteaching20-Research

13