CMSC848X · Fall 2026 · Computer Science · University of Maryland, College Park
| Instructor | Ruoshi Liu — ruoshi@umd.edu |
| Teaching assistant | George Zhang — qz2002@umd.edu |
| Meeting time | Tue & Thu, 14:00–15:15 |
| Location | CSI 1122 |
| Office hours | IRB 4218, Thu 15:30–16:30 |
| Credits | 3 |
| Discussion | Course Slack |
| Syllabus | Full syllabus (PDF) |
This course introduces the core methods that let robots learn from data. We focus on visuomotor policy learning for manipulation, covering behavior cloning, offline and online reinforcement learning, sim-to-real transfer, and multimodal perception across vision, touch, sound, and force. Alongside the methods, we take up the problems that make robotics distinctive — embodiment, data efficiency, and generalization across tasks and platforms — and the directions currently reshaping the field, including robotic foundation models and world models.
The course is run as a seminar built on recent research papers. There are no textbook lectures after the first week. Instead, students present, argue, and defend papers in structured roles, then discuss them in small groups. The semester ends with a research-style final project.
Graduate standing or instructor permission. Students should have prior coursework or equivalent experience in machine learning and deep learning, comfort with linear algebra and probability, and working fluency in Python and PyTorch. Prior robotics coursework is helpful but not required; prior experience with a physical robot is not assumed.
Tuesdays are Topic Days; Thursdays are Researcher Days. Weeks 1 and 15, and the two sessions stranded by Fall Break and Thanksgiving, are instructor lectures, guest lectures, and final presentations.
| Wk | Tuesday — Topic Day | Thursday — Researcher Day |
|---|---|---|
| 1 | Sep 1 · Instructor lecture Course Overview & Logistics What the seminar is, how the roles work, how to read a robotics paper, sign-ups. |
Sep 3 · Instructor lecture Foundations of Robot Learning Problem setup, embodiment, data, evaluation — the vocabulary for the rest of the term. |
| 2 | Sep 8 · Topic Day Visual Representation for Robotics DINOv1 · DINOv2 · DINOv3 |
Sep 10 · Researcher Day Jitendra Malik Group-selected papers |
| 3 | Sep 15 · Topic Day Behavior Cloning IBC · ACT |
Sep 17 · Researcher Day Shuran Song Group-selected papers |
| 4 | Sep 22 · Topic Day Reinforcement Learning DPPO · TD-MPC2 |
Sep 24 · Researcher Day Richard Sutton Group-selected papers |
| 5 | Sep 29 · Topic Day Simulation & Sim-to-Real Transfer SimToolReal · RMA |
Oct 1 · Researcher Day Xue Bin Peng Group-selected papers |
| 6 | Oct 6 · Topic Day World Models Curiosity · UniSim |
Oct 8 · Researcher Day Danijar Hafner Group-selected papers |
| 7 | Oct 13 Fall Break — no class |
Oct 15 · Guest lecture Huy Ha Anthropic |
| 8 | Oct 20 · Topic Day Language Models in Robotics π0.7 · VoxPoser |
Oct 22 · Researcher Day Andy Zeng Group-selected papers |
| 9 | Oct 27 · Topic Day Multisensory & Multimodal Perception See, Hear, and Feel · ManiWAV |
Oct 29 · Researcher Day Andrew Owens Group-selected papers |
| 10 | Nov 3 · Topic Day Tactile Manipulation Reactive Diffusion Policy · 3D-ViTac |
Nov 5 · Researcher Day Roberto Calandra Group-selected papers |
| 11 | Nov 10 · Topic Day Grasping Dexonomy · SPIDER |
Nov 12 · Researcher Day Ken Goldberg Group-selected papers |
| 12 | Nov 17 · Topic Day Dexterous Manipulation DextrAH-RGB · Visual Dexterity |
Nov 19 · Researcher Day Deepak Pathak Group-selected papers |
| 13 | Nov 24 · Guest lecture Daniel Seita USC |
Nov 26 Thanksgiving — no class |
| 14 | Dec 1 · Topic Day Computational Hardware Design DGDM · Hardware as Policy |
Dec 3 · Researcher Day Wojciech Matusik Group-selected papers |
| 15 | Dec 8 · Final presentations Final Project Presentations I |
Dec 10 · Final presentations Final Project Presentations II |
No final exam for this course.
The week has a fixed rhythm. Tuesdays are Topic Days and Thursdays are Researcher Days. Every session runs on the same 75-minute clock:
| Time | Segment | What happens |
|---|---|---|
| 45 min | Presentation | Student teams present, in role. Topic Days use the eight-role structure below; Researcher Days are group paper presentations. |
| 15 min | Breakout discussion | The class splits into groups of ~7 to discuss the papers away from the podium. |
| 15 min | Conclusion | Groups report back; the presenting team synthesizes and closes. |
Each Topic Day covers one theme through two or three papers. The presenting team opens with two framing slides — one on the topic as a whole, one situating the specific papers — and then splits into eight roles. Each role owns roughly one slide and five minutes.
The authors and reviewers are adversarial by design. The author's job is to sell the paper as hard as they honestly can. The reviewer's job is to reject it. Neither role is a summary; both are arguments. The learners, on the other hand, are neutral summarizers who try to discover high-level trends and ideas as well as low-level but crucial technical decisions.
| Role | Mandate |
|---|---|
| Task author | Frame the problem the paper attacks. Why does it matter, why is it hard, and what does solving it buy the field? Sell it. |
| Task reviewer | Argue the problem is ill-posed, already solved, narrower than claimed, or not worth the field's attention. |
| Approach author | Present the method and defend every design choice as the right one under the constraints. |
| Approach reviewer | Attack the method: unjustified choices, hidden assumptions, brittle components, complexity without payoff. |
| Result author | Present the evidence and argue it supports the paper's claims. Highlight the experiments that actually matter. |
| Result reviewer | Attack the evaluation: weak baselines, cherry-picked tasks, missing ablations, generalization claims the data cannot bear. |
| High-level learner | What are the high-level ideas? How do they differ from prior work? What does this open up for future work? |
| Low-level learner | What is the key trick that made it work? What smart ML / robotics / algorithmic techniques were used? What would you do differently? |
The deck also carries three closing slides: one synthesizing the week's Slack thread, one posing the breakout prompts, and one conclusion drawn from what the breakout groups reported. Total deck: 2 framing + 8 role + 3 closing slides.
Thursdays are built around a single researcher rather than a single topic. Each group picks one paper by that researcher and presents it. Three rules:
Group presentations follow the Task / Approach / Result spine from Topic Days, without the adversarial split, and close with one high-level takeaway and one low-level trick. Each group has roughly 8–10 minutes depending on class size. The point of the day is to see how one researcher's questions, methods, and taste evolve across a body of work — so tell us how your paper fits that arc.
Before every class, each student posts one substantive question about the assigned papers or researchers and responds to one question posted by someone else. Both are due by 10:00 pm the day before class. A good question is one you could not answer by just reading the abstract. Presenters can take inspiration from the Slack discussions.
The class splits into groups of roughly seven, re-shuffled periodically so you are not always arguing with the same people. Discuss any of: something new you learned from the paper; an insight you formed while reading it; an idea or research direction it inspired; a reflection worth sharing with the rest of the class. Each group designates a reporter who has two minutes in the conclusion segment.
The final project is a small piece of original research in robot learning, carried out in teams of 2–4. It should look like a workshop paper: a clear question, an honest experiment, and a result you can defend — including a negative one. Reproducing a paper is acceptable only if you extend it with a question the original did not answer.
Projects may be simulation-only. Access to physical hardware is not required. Compute is available through the university cluster.
| Milestone | Due | What to submit |
|---|---|---|
| Team formation + topic sketch | Thu, Sep 17 | One paragraph and a team roster, posted in Slack. Ungraded but required. |
| Project proposal (1–2 pages) | Thu, Oct 8 | Question, related work, proposed method, evaluation plan, risks and fallback. |
| Mid-project check-in (1 page) | Thu, Nov 5 | What works, what broke, what changed. Paired with a required 15-minute meeting. |
| Final presentation | Dec 8 & 10 | 12–15 min per team plus questions, in class. |
| Final report + code | Thu, Dec 10 | 4–8 pages in a standard conference format, plus a runnable repository. |
| Component | Weight | Notes |
|---|---|---|
| Slack discussion | 10% | One question and one reply before each class. Graded on engagement and substance, not volume. |
| Course presentations | 50% | Topic Day role performance (30%) + Researcher Day group presentation (20%). |
| Final project | 40% | Proposal 5% · check-in 5% · final presentation 15% · report and code 15%. |
Grade scale: A 90–100, B 80–90, C 70–80.
| Criterion | Weight | What excellent looks like |
|---|---|---|
| Question and motivation | 20% | The question is sharp, unresolved by prior work, and worth a semester. |
| Technical execution | 30% | Sound method, correct implementation, sensible baselines, controlled comparisons. |
| Evidence and analysis | 25% | Claims are supported by the experiments run. Ablations isolate the thing you say matters. Negative results reported honestly. |
| Communication | 20% | Report and talk are clear, well-figured, and appropriately scoped. |
| Reproducibility | 5% | Code runs from a clean checkout with documented setup. |