Mobile Robot Systems — CS 4331/5331 · Fall 2026

Why take this course

A policy that is brilliant in a notebook is useless on a robot if it can't meet the latency budget — and dangerous if it fails silently. This course lives on the seam where deployed autonomy actually succeeds or fails.

Real hardware

End on a real robot

Every project develops sim-first, then deploys to Jetson-powered ground robots or Crazyflie micro-drones and measures the sim-to-real gap. "Does it actually work?" becomes a number.

Two-way bridge

ML or systems — not both

The course levels each side up on the other: the modular autonomy pipeline (sensing, perception, planning, control) for systems folks; real-time, edge, and inference systems for ML folks.

Project-driven

Build, measure, publish

No final exam. The capstone is a semester project — make an autonomy capability efficient, then defend it with measurements in a workshop-style paper and talk.

Two cliffs labeled Robot Learning and Computing Systems are connected by a blue truss bridge labeled Mobile Robot Systems — co-design across the seam; a small robot crosses toward a sign reading efficient autonomy, proven on real hardware. — Each community optimizes its half; deployed autonomy lives on the seam between them — this course is the bridge.

At a glance

dated sessions over 15 weeks

robots — 4 Jetson ground robots + 3 Crazyflie drones

ready-made project directions (or propose your own)

final exams — final presentations instead

Grading

Project 35% (5331) Labs 30% Presentation 10% Critiques 10% Midterm 10% In-class 5%

Team model: the course project is required of CS 5331 students only, individually or in a team of up to 3; CS 4331 students do not complete a project — the one difference between the sections. Percentages above are the CS 5331 ledger; without the project the CS 4331 ledger renormalizes to Labs 46.2%, presentation 15.4%, critiques 15.4%, midterm 15.4%, in-class 7.7%.

What you'll build

Eight project directions — every one touches both sides of the bridge and ends with a hardware reality-check. Or propose your own.

Anytime perception on a budget

An early-exit / cascaded detector; map the accuracy–latency–energy frontier.

Visuomotor policy on a deadline

A sim-trained driving policy quantized and accelerated until it hits the control deadline on-device; task success vs. latency, sim vs. real.

Adaptive edge offloading

Split perception between robot and edge under replayed network traces; the split adapts as the link degrades.

LiDAR localization on a memory budget

2-D SLAM with a bounded-memory map; the accuracy vs. memory/latency frontier; real drift vs. sim.

Compress a big model for the edge

Distill / quantize a large vision model onto the Orin Nano within a frame-time budget; measure the accuracy cost.

Sim-to-real warmup study

How a short, cheap on-robot adaptation phase closes the sim-to-real gap — the reality-check is the core of the project.

Collaborative perception

Two robots — or a drone swarm — share detections over WiFi; accuracy gain vs. bandwidth and latency cost.

Autonomy-stack measurement study

Profile a ROS 2 pipeline on the Orin Nano under load; locate the efficiency bottlenecks.

Left, a wireframe simulator view of a robot with metrics accuracy 91 percent, latency 38 milliseconds, energy 4.1 watts; a blue deploy arrow for weeks 12 to 14 leads to a photo-style physical robot with unknown metrics; dashed lines converge on a box reading delta equals the sim-to-real gap, above a 15-week timeline. — Every project shares one arc: iterate sim-first all term, deploy once for real in Weeks 12–14 — and the measured Δ is what your paper reports.

Schedule

28 sessions, five arcs, every one dated (Tue/Thu 8:00–9:20 AM). Most sessions from Week 3 on end with a 20-minute student paper slot.

The modular autonomy pipeline drawn on a robot chassis: Sense, Perceive, Plan, and Act boxes with labeled data arrows and Arc-1 session numbers, plus a blue arrow labeled end-to-end policy bypassing the middle stages from camera to control. — Arc 1 builds the pipeline module by module — sessions L4–L8 — then weighs it against the end-to-end shortcut (L9). All of it runs on the robot's compute budget.

Arc 0 — Framing Week 1

Session	Date	Topic	Paper / milestones
L1	Tue Aug 25	Bridging robot learning and computing systems. The two communities; what efficient end-to-end autonomy demands.	Presentation sign-up opens
L2	Thu Aug 27	A tour of real end-to-end autonomous systems. AV stacks, drones, manipulators — where efficiency breaks.	Instructor models the paper slot

Arc 1 — Pillar I: How autonomy is built & learned Weeks 2–5

Session	Date	Topic	Paper / milestones
L3	Tue Sep 1	The modular autonomy pipeline. Sense → perceive → plan → act; where learning enters.
L4	Thu Sep 3	Sensing and state estimation. Cameras, LiDAR, IMU; KF/EKF; learned estimation.	Lab 1 out
L5	Tue Sep 8	Perception I — localization & mapping. SLAM fundamentals.
L6	Thu Sep 10	Perception II — detection, segmentation, tracking.	YOLO · first student slot
L7	Tue Sep 15	Planning and navigation. Motion planning; Nav2; learned planning.	Nav2 · group + topic due
L8	Thu Sep 17	Control, and learning for control. PID → MPC; imitation learning; RL in brief.	Diffusion Policy
L9	Tue Sep 22	End-to-end & foundation models vs. the modular stack. VLAs — what E2E buys and costs.	RT-2 · Lab 1 due

Arc 2 — Pillar II: How robots compute Weeks 5–8

Session	Date	Topic	Paper / milestones
L10	Thu Sep 24	Robot software systems. ROS 2, the compute graph, execution.	ROS 2
L11	Tue Sep 29	Real-time systems for robots. Deadlines, WCET, scheduling.	ROS 2 response-time · Lab 2 out
L12	Thu Oct 1	The robot compute platform. CPU/GPU/NPU; power budgets; profiling the course Jetson.	Proposal due
L13	Tue Oct 6	ML and inference systems. How DNNs execute; runtimes and accelerators.	TVM
L14	Thu Oct 8	Edge, cloud, and connected robots. Offloading trade-space; V2X connectivity.	Neurosurgeon
L15	Tue Oct 13	Midterm exam — in class, open-book, covering both pillars.	Lab 2 due

Arc 3 — The Bridge: Efficiency Weeks 8–13

Session	Date	Topic	Paper / milestones
L16	Thu Oct 15	The efficiency gap. A learned policy vs. the robot's budget; how to measure.
L17	Tue Oct 20	Making models fit. Quantization, pruning, distillation.	Integer-only quantization · baseline due
L18	Thu Oct 22	Anytime and input-adaptive inference. Cascades, early exit; the accuracy–latency–energy frontier.	BranchyNet · Lab 3 out
L19	Tue Oct 27	System-level efficiency I — scheduling. Learned workloads under deadlines.	D3
L20	Thu Oct 29	Project milestone presentations.	Milestone
L21	Tue Nov 3	Hardware-aware model design & NAS. Efficient architectures; once-for-all networks.	Once-for-All
L22	Thu Nov 5	Energy and power management. Measuring energy on the Jetson; DVFS; energy-aware inference.	Energy-aware inference paper
L23	Tue Nov 10	System-level efficiency II — offloading & split computing. Robot↔edge splits on replayed traces; co-design.	Split-computing paper
L24	Thu Nov 12	Data systems for robot learning. Datasets, auto-labeling, the fleet data flywheel.	RT-X · Lab 3 due · reality-check opens
L25	Tue Nov 17	Efficient training & on-robot adaptation. Training vs. inference cost; PEFT; continual adaptation.	Efficient-training paper

Arc 4 — The end-to-end system Weeks 13–15 + finals

Session	Date	Topic	Paper / milestones
L26	Thu Nov 19	Integrating the system. Cross-layer co-design; sim to physical robot.
L27	Tue Nov 24	Connected and collaborative autonomy. V2X, cooperative perception, swarms, connected-vehicle stacks.	V2VNet
L28	Tue Dec 1	Final project presentations I.	Part II in the final-exam block (Dec 4–9) · report due Dec 9

Three labs, one ladder

Each lab builds a skill on one side of the bridge — the last crosses it on real hardware.

Stand up the pipeline in simulation

Bring up a TurtleBot 4 or a Clearpath Husky in Gazebo, map its world with SLAM, and navigate the map with Nav2 — sense → perceive → plan → act. Out Sep 3 · due Sep 22.

Profile it as a system

Instrument that navigation stack — a ROS 2 graph — and measure end-to-end latency and where time actually goes. Out Sep 29 · due Oct 13.

Cross the bridge on real hardware

Deploy to a Jetson robot: compress and accelerate until it hits an on-device latency budget — measure accuracy, latency, and energy on the device. Out Oct 22 · due Nov 12.

The fleet

🚗 Jetson ground robots ×4

JetBot-class differential drive: NVIDIA Orin Nano + camera + 2-D LiDAR. The on-device efficiency platform — compression, anytime inference, and localization projects run here.

🚁 Crazyflie drones ×3

27-gram open-source micro-quadcopters — too small to compute onboard, so they embody the offloading problem, have a sharper sim-to-real gap, and are swarm-native.

Development is sim-first (Gazebo / Isaac Sim, public datasets, cloud GPUs); the fleet is booked late-term for every project's hardware reality-check.

Register & logistics

How to register

Register on Raiderlink for Fall 2026: CS 4331 (senior undergraduates) or CS 5331 (graduate students) — dual-listed, meeting together Tue/Thu 8:00–9:20 AM in Holden Hall 00075. Questions about fit or background? Email liangkai.liu@ttu.edu — or drop by office hours, Tue 2–3 PM & by appointment.

What you need coming in

Python and C++, linear algebra and probability, and a background in either ML or systems — the course levels you up on the other side. No robotics experience assumed. CS 4331 expects CS 2413 plus one of CS 3375 / CS 4352 / an intro AI-ML course.