|Title||CISC849 Robot Vision and Learning|
|Description||Survey of image-based 2-D and 3-D sensing algorithms for mobile robot navigation and interaction, including motion estimation, obstacle segmentation, terrain modeling, and object recognition, with a particular focus on deep learning techniques to dramatically improve performance.|
|When||Tuesdays and Thursdays, 2-3:15 pm|
|Instructor||Christopher Rasmussen, 446 Smith Hall, email@example.com|
|Office hours||Mondays and Wednesdays, 2-3 pm|
|Academic policies||Programming projects are due by midnight of the deadline day (with a grace period of a few hours afterward...after sunrise is definitely late). A late homework is a 0 without a valid prior excuse. To give you a little flexibility, you have 6 "late days" to use on homeworks to extend the deadline by one day each without penalty. No more than three late days may be used per assignment. Late days will automatically be subtracted, but as a courtesy please notify the instructor in an e-mail of your intention to use late days before the deadline. See submission instructions below.
Students can discuss problems with one another in general terms, but must work independently on programming assignments. This also applies to online and printed resources: you may consult them as references (as long as you cite them), but the words and source code you turn in must be yours alone. The University's policies on academic dishonesty are set forth in the student code of conduct here.
|Homeworks||Assignment submissions should consist of a directory containing all code (your .cpp files, makefile, etc.), any output data generated (e.g., images, movies, etc.), and an explanation of your approach, what worked and didn't work, etc. contained in a separate text or HTML file. Do not submit executables or .o files, please! The directory you submit for each assignment should be packaged by tar'ing and gzip'ing it or just zip'ing it. The resulting file should be submitted through Canvas.
You may develop your C/C++ code in any fashion that is convenient--that is, with any compiler and operating system that you want. However, we will be grading your homework on a Linux system with a makefile, and so you must avoid OS- and hardware-specific functions and provide a makefile for us that will work (like one of the templates above).
Possible Papers to Present (not a complete list)
- Where Should I Walk? Predicting Terrain Properties from Images via Self-Supervised Learning, Wellhausen et al., IEEE Robotics and Automation Letters, 2019
- Supersizing Self-supervision: Learning to Grasp from 50K Tries and 700 Robot Hours, Pinto and Gupta, ICRA 2016
- Learning real manipulation tasks from virtual demonstrations using LSTM, Rahmatizadeh et al., AAAI 2018
- Learning to Fly by Crashing, Gandhi, Pinto, and Gupta, IROS 2017
- VoxNet: A 3D Convolutional Neural Network for real-time object recognition, Maturana and Scherer, IROS 2015
- PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation, Qi, Su, Mo, and Guibas, CVPR 2017
- Real-Time Navigation in 3D Environments Based on Depth Camera Data, D. Maier, A. Hornung, and M. Bennewitz, Humanoids 2012
- Robotic Grasping of Novel Objects using Vision, A. Saxena, J. Driemeyer, A. Ng, IJRR 2008
- DeepLoco: Dynamic Locomotion Skills Using Hierarchical Deep Reinforcement Learning, X. Peng, G. Berseth, K. Yin, and M. van de Panne, SIGGRAPH 2017
- High fidelity day/night stereo mapping with vegetation and negative obstacle detection for vision-in-the-loop walking, M. Bajracharya et al., IROS 2013
Note: The blue squares in the "#" column below indicate Tuesdays.
|2||Aug. 29||Introduction to DARPA Urban Challenge (UC), Robotics Challenge (DRC) and algorithm components||
|3||Sep. 3||More about Grand Challenges, trail-following||slides|
|4||Sep. 5||Traditional 2-D/3-D shape/object segmentation, part I|
|6||Sep. 12||Traditional 2-D/3-D shape/object segmentation, part II||slides|
|7||Sep. 17||Image classification background||Stanford CS231n Image classification slides (6-61),|
|8||Sep. 19||NO CLASS|
|9||Sep. 24||More classification background; Introduction to convolutional neural networks||Stanford CS231n: Loss/optimization slides (8-10, 37-77, finish with web demo), Intro to NN (83-88), Convolutional Neural Networks slides (15-63, 71-78), CNN architectures (8-36)|
|10||Sep. 26||Detection & segmentation background||YOLOv2, Redmon and Farhadi (CVPR 2017)||Stanford CS231n Detection and Segmentation slides (15-27, 44-85, 89-94)|
|11||Oct. 1||TensorFlow programming and homework||Deep learning software slides (21-33, 40-59)|
|Oct. 3||NO CLASS
|Oct. 8||NO CLASS
|12||Oct. 10||Perception for stepping||"Learning Locomotion over Rough Terrain using Terrain Templates", M. Kalakrishnan, J. Buchli, P. Pastor, and S. Schaal, IROS 2009||ANYmal video, ANYmal slides, ANYmal "Where Should I Walk?" video, Kalakrishnan slides|
|13||Oct. 15||NO CLASS
|14||Oct. 17||Imitation learning||
||Berkeley Levine course IL slides (DAgger defined on slide 14); CMU Fragkiadaki IL slides (1-29)
|Reinforcement learning||UCL Silver course: RL slides (7-42), RL video (15:40-21:20, 1:02:13-1:04:40 contains movies missing from PDF); Stanford deep-Q lecture slides (23-62); original deep-Q paper|
|16||Oct. 24||NO CLASS
Break before presentations start
|17||Oct. 29||Student paper presentation||Maxim Bazik, "PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud"|
|18||Oct. 31||Student paper presentation||Woosik Lee, "Learning Wheel Odometry and IMU Errors for Localization"|
|19||Nov. 5||Student paper presentations||Yi Liu, "Reinforced Cross-Modal Matching and Self-Supervised Imitation Learning for Vision-Language Navigation"|
|20||Nov. 7||Student paper presentations||Chuchu Chen, "Deep Learning for 2D Scan Matching and Loop Closure"
Chirantan Ghosh, "Video Action Transformer Network"
|21||Nov. 12||Student paper presentation||Sha Liu, "Robots that can adapt like animals"
Seyedalireza Khoshsirat, "PCN: Point Completion Network"
|22||Nov. 14||Student paper presentations||Shengye Li, "Learning Common and Specific Features for RGB-D Semantic Segmentaion with Deconvolutional Networks"
Eric Wright, "Generalization through Simulation: Integrating Simulated and Real Data into Deep Reinforcement Learning for Vision-Based Autonomous Flight"
|23||Nov. 19||Student paper presentation||Shivanand Sheshappanavar, "PointFlow: 3D Point Cloud Generation with Continuous Normalizing Flows"
Zhang Guo, "Believe It or Not, We Know What You Are Looking at!"
|24||Nov. 21||Student paper presentations||Nikhil Gothankar, "Fast Online Object Tracking and Segmentation: A Unifying Approach"
Vineet Singh, "Spatial Transformer Networks"
|Nov. 26||NO CLASS
|Nov. 28||NO CLASS
|25||Dec. 3||Student paper presentation||Md Mottalib, "An Empirical Evaluation of Deep Learning on Highway Driving"|
|26||Dec. 5, 6||Final project presentations|