Where is my robot butler?

  • Why is everyone using language models?
  • Why use learning based methods when control works?
  • What's with Foundation models for robotics?
  • Will GPT-5 be a robot? 😱
Household robots need to move beyond simple programmed tasks like those a Roomba performs and become full-fledged digital assistants.

Bring your own robot, let's add language to your research

A robotic agent that exists (physically) in the world, gains access to rich and personalized knowledge of its environment. For example, they might be able to answer questions like: How much do things weigh? What's fragile? Or where you store the extra chocolates that you don't want anyone to find? Building an agent that can accomplish tasks requires the integration of a diverse set of technologies and engineering. Language models, SLAM, semantic mapping, task planning, understanding object affordances, and end effector control. This course will cover both foundational works in grounding language to action and analyze (or reimplement) state-of-the-art Large Language Model based task planners. This area is fascinating and difficult because it is so cross-cutting. Readings and topics are pulled from Robotics (CoRL, ICRA, RSS, HRI, IROS), Computer Vision (CVPR, I/ECCV), Natural Language Processing (ACL, EMNLP), and Machine Learning (ICLR, ICML, NeurIPS).

Projects will be scoped by prior hardware/simulator experience -- but knowledge of Deep Learning + one specialty (NLP/CV/Robotics) is basically required. Send Qs to Yonatan (ybisk@cs).

Topics
  • LLMs & Foundation Models
  • Instruction following & Dialogue
  • Task and Motion Planning
  • End-Effector & real-valued control
  • Semantic Mapping (2D and 3D)
  • World Models
Questions
  • How do you define or evaluate Dialogue?
  • Limitations of offline and unimodal pretraining
  • How does embodiment shape meaning?
  • Discrete vs continuous spaces and representations.
  • When is Sim2Real possible? What's about manipulation?
  • I only have one brain, do I need more than one model?
Logistics
A basic course schedule is presented below, additional readings here: Additional Readings
Tues
Thurs

What is embodiment? Designing your world

Aug 27: Philosophy
Aug 29: Origins of the Field
Sept 3: Simulators, Action Spaces, Robotics
Sept 5: NLP, LLMs, VLMs, ...
  • Syntactic/Semantic Parsing
  • (Large) Language Models
  • Multimodal Transformers
Sept 10: No class (Prep Project Pitch)
Sept 12: Initial Group Pitches
Sept 17: Perspective Excercise (Argue merits from the perspective of another field of ML)
Sept 19: Red team your project

Navigation

Sept 24: Discrete Worlds
Sept 26: Continuous Worlds
Project Report: Assumpetions about meaning and Linguistic Analysis
Oct 1: Mapping and Replanning
Oct 3: Real Valued output and Sim2Real
Oct 8: State Tracking and Task Planning
Oct 10: Language Models as Planners
Project Report: Code-as-policies, Progprompt, ...

Manipulation

Fall Break
Oct 22: Large _____ Models
Oct 24: Manipulators & Representing Space
Oct 29: Imagination
Oct 31: Mid-Term Presentations (+Feedback)

Feedback, Dialogue, and Teaching

Nov 5: Election Day Nov 7: Guest Lecture (e.g. Daniel Fried)
Nov 12: Concept Learning
Nov 14: Project Office Hours (virtual) Sign-up sheet in canvas
Nov 19: Feedback
Nov 21: Asking for Help and Dialogue
Nov 26: Theory-of-mind
Nov 28: Thanksgiving, no class
Dec 3: Revisit Philosophy and Open Challenges Dec 5: Final Presentations
Here are a few things that haven't made it into the schedule yet
This course is available as both a seminar (6 credits) and project based (12 credits) course.
Grades
Paper Summaries See format below 5pts * 8 papers (indiv)
Participation/Pres.
Sept 12 Project Pitch 5pts (group)
Sept 17 Perspective exercise 5pts (indiv)
Sept 19 Red Teaming 5pts (indiv)
Oct 31 Midterm Presentation 5 pts (group)
Oct 31 Midterm feedback 5 pts (indiv)
Dec 5 Final presentation 10 pts (group)
Project Reports
Sept 26 Meaning & Ling Analysis 5 pts (group)
Oct 10 Code-as-policies, ... 5 pts (group)
Nov 14 Pragmatics and Legibility 5 pts (group)
Dec 5 Final Report 10 pts (group)
Proposal: Both seminar and project based students will write a proposal. While project students will go on to work on implementation, the seminar students should also go through the mental exercise of planning out what a system needs, what dependencies components have, where gradients might flow, etc. They will then get to revise their understanding in their final report.

Groups: Both seminar and project based assignments will be done in groups. Groups will likely be capped at five people.

Equal Participation: All reports must include a breakdown of each teammate's contributions.

Paper Summary (5pts)
  1. What assumptions were made about:
    • meaning
    • the environment
    • interaction / social intelligence
  2. What extension would you propose? Scope it to less than 9mos of work, and justify
  3. What do you really like about this work?
Project Pitch (5pts)
  • Task, Environment, and Skills Definitions
  • Minimal language covered and stretch goals
  • Who do you think will be most interested?
  • What will your project enable in the future?
  • Paint me a picture of how/where/who will use this
Midsemester Presentation (5pts)
  • Interactive demo of basic skills
  • Example of successful composition
  • Demonstration and analysis of failures
  • Proposal of changes for final demo (including rescoping)
Final Presentation (10pts)
  • Remind everyone of what you're doing with demo of compositional instructions
  • Updated technical hypothesis if you were to start over
  • Evidence (e.g. from demonstration or analysis) for your new hypothesis
Final Report (10pts)
  • 12 Credit: Technical write-up and specification of system
  • 12 Credit: Technical write-up of model design
  • All: Updated design elements for the intermediate reports
  • All: Literature Review of state of the field
  • All: Discussion of key limitations to progress in this space
  • All: Team me something or try and change my mind about something from class
The course will be primarily centered on a few robot platforms or simulators based on the size of the course enrollment and the prior experience of students taking the class for 12 credits.

There are a couple other simulators/platforms I also like, which we can discuss as options.

Platform Type Notes
VLN-CE Simulated Navigation
Minimal hardware experience
Proj: Language to Angle/Distance
Teams:No limit on teams
DexArm Simple gripper
Basic manipulation platform
Proj: Language to 3D Space
Teams: Two teams of ~4
Hello Robot Stretch Mobile Manipulator
Requires skill specifications
Proj: Language to ... let's decide
Teams (Probably) one team -- Let's see
Control Code: Meta Home-Robot
Collaborator: Chris Paxton
Late Assignments
  • All teams have 5 late days, these are only applicable to reports (not demos).
  • Paper summaries lose 1pt per day late

COVID Details:

In the event a student tests positive for COVID-19, they will be invited to attend discussion virtually and will be expected to participate as usual. This includes participation points for raising their hands with questions/answers and submission of lab-notebooks. Note, that students who attend class while exhibiting symptoms will be told to leave and join virtually for the protection of all others present.

Accommodations for Students with Disabilities:

If you have a disability and have an accommodations letter from the Disability Resources office, we encourage you to discuss your accommodations and needs with the instructors as early in the semester as possible. We will work with you to ensure that accommodations are provided as appropriate. If you suspect that you may have a disability and would benefit from accommodations but are not yet registered with the Office of Disability Resources, we encourage you to contact them at access@andrew.cmu.edu.

  1. Can we use other platforms? Yes! What robots do you have? Also checkout AI Maker Space
  2. What about custom sensors and hardware? Same answer :)
  3. What about other simulators? Same answer :)
  4. LTI Curriculum Categories? 12 Hour version can be counted for a Task and a Lab
  5. Do I /need/ simulator experience? No, but plan to spend some time getting the engineering setup
  6. Can I attend discussion without registering? It's best to register (6hrs) even if you've finished your classes, since I need to prioritize time, energy, and space on registered students. I'll try and update this once I have a room confirmed with the registrar and see how much space we have in the class.