COMP34212 Cognitive Robotics

COMP34212 Cognitive Robotics

Angelo Cangelosi

COMP34212: Coursework on Deep Learning and Robotics

34212-Lab-S-Report

Angelo Cangelosi<angelo.cangelosi@manchester.ac.uk>

Release: February 2025Submission deadline: 27 March 2025, 18:00 (BlackBoard)

Aim and Deliverable

The aim of this coursework is (i) to analyse the role of the deep learning approach within the

ontext of the state of the art in robotics, and (ii) to develop skills on the design, execution and

evaluation of deep neural networks experiments for a vision recognition task. The assignment willin particular address the learning outcome LO1 on theanalysis of the methods and softwaretechnologies for robotics, and LO3 on applying different machine learning methods for intelligentbehaviour.

The first task is to do a brief literature review of deep learning models in robotics. You can give asummary discussion of various applications of DNN to different robotics domains/applications.

Alternatively, you can focus on one robotic application, and discuss the different DNN models used

for this application. In either case, the report should show a good understanding of the key works inthe topic chosen.The second task is to extend the deep learning laboratory exercises (e.g. Multi-Layer Perceptron(MLP) and/or Convolutional Neural Network (CNN) exercises for image datasets) and carry outandanalyse new training simulations. This will allow you to evaluate the role of differenthyperparameter values and explain and interpret the general pattern of results to optimise thetraining for robotics (vision) applications.You can use the standard object recognition datasets (e.g. CIFAR, COCO, not the simple MNIST) orrobotics vision datasets (e.g. iCub 代写COMP34212 Cognitive RoboticsWorld1 , RGB-D Object Dataset2 ). You are also allowed to useother deep learning models beyond those presented in the lab.The deliverable to submit is a report (max 5 pages including figures/tables and references) to

describe and discuss the training simulations done and their context within robotics research andappendix (the Code Appendix is in addition to the 5 pages of the core report). Do not use AI/LLMmodels to generate your report. Demonstrate a credible analysis and discussion of your ownimulation setup andresults, not of generic CNN simulations. And demonstrate a credible,

ersonalised analysis of the literature backed by cited references.

Marking Criteria (out of 30)

1. Contextualisation and state of the art in robotics and deep learning, with proper use ocitations backing your academic review and statements (marks given forclarity/completeness of the overview of the state of the art, with spectrum of deep learninmethods considered in robotics; credible personalised critical analysis of the deep learningrole in robotics; quality and use of the references cited) [10]

2. A clear introductory to the DNN classification problem and the methodology used, witexplanation and justification of the dataset, the network topology and the hyperparameterschosen; Add Link to the code/notebook you used or add the code in appendix. [3]Complexity of the network(s), hyperparameters and dataset (marks given for complexityand appropriateness of the network topology; hyperparameter exploration approach; dataprocessing and coding requirements) [4]

4. Description, interpretation, and assessment of the results on the hyperparameter testingsimulations; include appropriate figures and tables to support the results; depth of theinterpretation and assessment of the quality of the results (the text must clearly andcredibly explain the data in the /tables); Discussion of alternative/future simulationsto complement the results obtained) [13]10% Marks lost if report longer than the required maximum of 5 pages: 10% Marks lost ifcode/notebook (link to external repository or as appendix) is not included.

Due Date: 27 March 2025, 18:00, pdf on Blackboard. Use standard file name: 34212-Lab-S-Report