UC Berkeley SERL: Learning with Less Practice Data
First understand what UC Berkeley researchers are trying to learn through RAIL, why the question matters, how they test it, and what the result cannot prove. Give a simulated robot position-and-speed numbers, camera images, or human demonstrations and compare which helps it learn faster. The student then completes A robot-learning management app that tracks learning efficiency and requests demonstrations when needed. For a college application or interview, the student separates the source research from their feature, interface, tests, failures, and revisions. Available online or in person in Apgujeong, in one-to-one or small-group formats.
University and lab names show where each public project came from. CIT independently designed these courses; they are not official university courses, partnerships, or endorsements.
8guided sessions
6-12 hoursto the first small project
Intermediate to advancedrecommended level
Not required for franka_simphysical robot
Primary routes
Choose the student's primary AI education and portfolio route
CIT keeps the international-school route first, followed by the separate science and gifted-school route.
CIT lessons can run online across Korea and overseas or in person in Apgujeong. The student owns the question, code, tests and explanation; classes do not guarantee admission, selection or awards, and online delivery never changes an institution's submission rules.
Which information and demonstrations help the same robot learn with fewer practice attempts?
Teach the same simulated robot task with position-and-speed numbers, camera images, and human demonstrations. Compare which input reduces the practice needed.
Students first use a small working example to understand the researchers' question, method, and evidence. A controlled change helps identify what the student's extension must solve. The student then implements a useful feature, connects an operator interface, and tests the integrated application in normal, boundary, and failure cases.
What will the student complete?
A robot-learning management app that tracks learning efficiency and requests demonstrations when needed
The final package includes runnable instructions, the feature and interface design, normal and failure tests, one documented revision, and a three-minute explanation in the student's own words.
Feature
A feature that selects useful human demonstrations and requests another example when learning stalls or repeatedly fails
Operator interface
A learning-progress screen showing practice count, success rate, demonstration timing, and representative failure clips
Optional LLM boundary
Optional: an LLM reads only student run records and failure notes to draft questions for the next demonstration. It does not choose robot actions.
Integrated result
A robot-learning management app that tracks learning efficiency and requests demonstrations when needed
Four ideas explained in this course
01learning with fewer attempts
02robot state as numbers
03scenes viewed by camera
04human demonstrations
Eight introductory sessions
Understand the research, then build and test a working robotics application
The 20 courses are eight-session CIT studios in which students understand a public university or lab project, confirm a working example, and then turn it into a small robotics application with a useful feature and an operator interface. Selected courses add an optional LLM explanation tool that can read run records but cannot control the robot. RoboMaster has 28 sessions, and implementation scope is adjusted to each student's experience and computer access.
Understand the source
01
Run the prepared simulated robot arm
02
Understand one practice attempt and its score
Measure and compare
03
Learn first from number data only
04
Learn from camera images
Build a feature
05
Change only whether human demonstrations are used
06
Implement the student-owned feature: A feature that selects useful human demonstrations and requests another example when learning stalls or repeatedly fails
Integrate and demonstrate
07
Build the operator interface: A learning-progress screen showing practice count, success rate, demonstration timing, and representative failure clips. Add the bounded assistant: Optional: an LLM reads only student run records and failure notes to draft questions for the next demonstration. It does not choose robot actions.
08
Integrate, test, and demonstrate: A robot-learning management app that tracks learning efficiency and requests demonstrations when needed
Public project used in class
See the exact version CIT reviewed
Public projects can change over time. To keep the class example consistent, CIT uses version 1fa2af7496be of rail-berkeley/serl. CIT checked it on 2026-08-14; it was created on 2025-10-27. Usage-rights note: Apache-2.0.
The student separates the parts of a robot that learns by trying, and compares runs with and without the demonstration data. In class the project is separated into five boxes (input, memory, process, output, control). The lesson opens one of them, changes one value, and leaves the rest closed.
InputWhat comes in?
Wrist-camera images, robot joint states, and the human demonstrations recorded up front
MemoryWhat persists?
The replay buffer holding past attempts, and the network weights that learning changes
ProcessWhat transforms?
The classifier that judges success, and the learning step that updates the policy from that judgement
OutputWhat leaves, and who uses it?
The next motion command sent to the robot server, and the training log
Controlopened hereWhat decides when anything runs?
The launcher that starts the acting side and the learning side separately and sets the order between them
The files the lesson opens, by name. The course is not a walk through the repository; it opens a chosen few and says which.
One value moves and everything else stays. The right-hand column is the prediction written before the run, not a result; the work is reconciling the two.
The universities and labs named here made the open projects this course reads. CIT designed the course independently; it is not an official, affiliated, or endorsed course.
Clear answers about what students do, what they need, and where the project came from.
Is UC Berkeley SERL: Learning with Less Practice Data an official course from UC Berkeley RAIL?
University and lab names show where each public project came from. CIT independently designed these courses; they are not official university courses, partnerships, or endorsements.
Is hardware required?
Not required for franka_sim. The course starts with a robot on the computer or a saved recording of a completed run. Computer guidance: GPU recommended for training.
What background should a student have?
Recommended level: Intermediate to advanced. Students should be able to follow a guided Python example, test one controlled change, and then build and explain a small feature and interface.
What will the student make?
The completed project is A robot-learning management app that tracks learning efficiency and requests demonstrations when needed. The student implements A feature that selects useful human demonstrations and requests another example when learning stalls or repeatedly fails and A learning-progress screen showing practice count, success rate, demonstration timing, and representative failure clips. Optional: an LLM reads only student run records and failure notes to draft questions for the next demonstration. It does not choose robot actions.
How can the student use this project in a college application?
Explain the source research question, method, evidence, and limits first. Then separate the student's own feature and interface decisions, normal and failure tests, revisions, and next question. The source institution's name does not imply affiliation or guarantee admission.
Which project version does the course use?
CIT reviewed version 1fa2af7496be of rail-berkeley/serl on 2026-08-14. That version was created on 2025-10-27. We keep this version during class so the example does not change unexpectedly, and we check the setup again before teaching.
Can UC Berkeley SERL: Learning with Less Practice Data be taken online or in person, one-to-one or in a small group?
Yes. This course is offered online and in person at CIT in Apgujeong, Gangnam-gu, Seoul, with one-to-one and small-group options. Placement and current availability are confirmed after a readiness consultation.
Is this a good first project for this student?
Before placement, we check the student's coding and math experience, available computer, interests, and ability to explain what happened.
UCSD ManiSkill: Teaching Many Simulated Robots at OnceTrain many simulated robots at the same time and compare how object size, friction, and goal position affect...→
Stanford robomimic: Learning Robot Actions from DemonstrationsReplay robot actions demonstrated by people, then compare how the number of examples and camera information...→
Georgia Tech MimicLabs: Creating Robot Practice DataTurn a small number of robot demonstrations into many practice scenes, then check whether the larger set adds...→