Source-linked AI summary
Teaching Reinforcement Learning and Humanoid Robotics to High-School Students: An Expert-Validated Curriculum Design on a Low-Cost Open Platform
Yuanzhe Dong, Jie Cao, Shuman Wang
TL;DR
Adapting complete robotics and RL workflows for high-school learners requires prerequisite support, manageable pacing, meaningful choices, and individual assessment. The framework organizes work around one robot, using staged tasks and bounded customization, and was refined through formative review by five experts. Students completed the course and achieved robot walking; expert ratings placed integration and teamwork highest (both M = 4.60, SD = 0.89).
Problem
Adapting complete robotics and RL workflows for high-school learners requires prerequisite support, manageable pacing, meaningful choices, and individual assessment.
Source: §I · pp. 1–2Method
The framework organizes work around one robot, using staged tasks and bounded customization, and was refined through formative review by five experts.
Source: §I · p. 2, B. Design Principles · p. 2Results
Students completed the course and achieved robot walking; expert ratings placed integration and teamwork highest (both M = 4.60, SD = 0.89).
Source: §I · p. 2, B. Quantitative Findings · p. 5Takeaways & Limitations
The framework adapts an assembly-to-deployment workflow for high-school students through staged tasks and bounded customization.
Source: §I · p. 2Five purposively selected experts provide formative feedback, not representative agreement, and the ratings establish neither formal content validity nor learning effectiveness.
Source: §VI · p. 6
Abstract
from arXiv · showhide
Lower cost open source robots and reinforcement learning (RL) simulation tools create new opportunities for precollege students to engage with contemporary robotics. However, translating a complete research workflow, spanning mechanical assembly, electrical setup, simulation, policy learning, system identification, and physical deployment, into a coherent course for novice learners remains challenging. We present an integrated robotics course framework that organizes these activities around a shared robotic artifact. The framework combines parallel disciplinary tracks, sequencing based on technical dependencies, progressive integration of simulation and hardware, layered performance checkpoints, and structures for balancing collaborative work with individual accountability. We illustrate the framework through a high school curriculum organized around a robot project in which pairs of students assemble an open source humanoid robot, train a walking policy in simulation, and deploy it on the physical platform. The framework was developed through an iterative design process that included formative review by five experts in robotics research, engineering, secondary STEM education, and curriculum design. Expert feedback highlighted three central design tensions: authenticity versus cognitive load, system integration versus timely visible progress, and team construction versus individual accountability. These tensions informed the final framework presented in this paper. This work offers a structured approach for adapting robotics research workflows into interdisciplinary precollege courses; future classroom studies are needed to examine implementation and student learning.
I. INTRODUCTION
The paper addresses the challenge of adapting complete robotics and reinforcement-learning workflows for high-school learners. It positions its framework as a way to connect prior RL activities to a student-assembled humanoid workflow while supporting prerequisite learning, pacing, choices, and assessment.
I. INTRODUCTION
- Teaching reinforcement learning introduces hardware costs, space constraints, and the distinction between programming and training behavior.Calls for deeper learner participation in RL design motivate connecting introductory activities to a complete robot-learning workflow.Source: §I · p. 1
- Adapting open humanoids and simulation tools for high school requires prerequisite support, manageable pacing, meaningful choices, and individual assessment.Open humanoids include ToddlerBot and Berkeley Humanoid Lite, while NVIDIA Isaac Lab supports simulation, training, and testing.Source: §I · p. 2
- The framework adapts assembly-to-deployment for high-school students through staged tasks and bounded customization, making model adjustment and physical testing explicit learning objectives.The paper reports course completion and robot walking, while reserving student performance and learning data for a separate study.Source: §I · p. 2
II. RELATED WORK
- Prior pre-college work already combines RL with virtual or physical robots and learner customization; this paper focuses on following a policy through assembly, measurement, simulation training, deployment, and individual explanations.The cited examples include ARtonomous, Zhang et al., and Balancing Act.Source: §II · p. 2
- The adaptation problem is to reduce prerequisite demands while preserving meaningful decisions and assessable understanding across the workflow.The paper extends the scope of learning opportunities but does not establish that a humanoid course teaches introductory RL more effectively than simpler alternatives.Source: §II · p. 2
III. COURSE DESIGN FRAMEWORK
The educational rationale connects multiple STEM disciplines through a shared, meaningful context. Scaffolding and formative assessment motivate the design, but its effectiveness remains untested.
A. Educational Rationale
- The shared robot connects prediction, measurement, and explanation within an integrated STEM context.Scaffolding limits unfamiliar concepts and tools, while formative assessment uses explanations to guide instruction; effectiveness remains untested.Source: A. Educational Rationale · p. 2
B. Design Principles
The framework coordinates work around one robot, sequences tasks by technical dependencies, and progressively integrates simulation with hardware. It pairs scaffolded checkpoints and bounded choices with individual evidence, while proposing controlled comparisons to support RL understanding.
B. Design Principles
- Coordinate mechanical, electrical, and computational work through observable behavior of the same robot.Students can explain a failed standing attempt using both balance and control concepts.Source: B. Design Principles · p. 2
- Sequence safety induction, component checks, power-on, standing, and walking as observable gates, while simulation can proceed before hardware is ready.This provides early conceptual progress while assembly continues.Source: B. Design Principles · p. 2
- Move from simulated predictions to physical measurements, testing a supplied policy before a student-trained policy to distinguish equipment faults, simulation errors, and learning failures..Source: B. Design Principles · p. 2
- Progress from worked examples to controlled modifications and independent decisions, with students justifying choices within instructor-set limits.Students interpret graphs and compare behavior; independent algorithm implementation is an extension.Source: B. Design Principles · p. 2
- Rotate pair roles so both learners predict, test, and explain, while recording individual reasoning separately from team robot performance.Individual reasoning can guide support before a harder task is assigned.Source: B. Design Principles · p. 2
C. Intended Support for RL Understanding
- Comparing one controlled change with a baseline connects reward design, generalization, and model limitations across virtual and physical trials.These are proposed learning mechanisms, not measured gains.Source: C. Intended Support for RL Understanding · p. 2
IV. HUMANOID CURRICULUM INSTANTIATION
The curriculum pairs high-school students around ToddlerBot and uses individual explanations alongside team construction. Its intended outcomes span safe assembly, RL concepts, simulation-to-physical comparisons, diagnosis, and collaboration.
A. Implementation Model
- The curriculum targets grades 9–12, with four students in two pairs, one ToddlerBot per pair, and a recommended mentor-to-student ratio of 1:4.Expert concerns prompted a prerequisite revision: students should run and modify a short Python program before RL labs; beginners need an introductory activity and additional preparation time.Source: A. Implementation Model · p. 2
- Eight three-hour CS+AI sessions provide 24 contact hours alongside assembly, electrical checks, and a closing demonstration.Additional build, preparation, and training time is not quantified.Source: A. Implementation Model · pp. 2–3
- Intended outcomes include safe assembly, explaining rewards and learned policies, comparing simulation with physical measurements, diagnosing differences, and explaining design decisions.Independent algorithm implementation and successful student-trained walking are extensions.Source: A. Implementation Model · p. 3
- Diagnosing a walking failure connects assembly checks, reward choices, and simulation comparisons; individual explanations support the collaboration outcome.Table IV specifies evidence for each outcome.Source: A. Implementation Model · p. 3
B. Platform
The platform pairs open hardware and software with bounded customization, while four coordinated tracks connect disciplinary ideas to practical robot decisions. Hardware work follows safety and testing gates, but simulation can provide early opportunities while construction continues.
B. Platform
- ToddlerBot provides open hardware and software, with reported parts cost under USD 6,000; fabrication, repairs, computing, and mentoring add delivery costs.Shared computers support training, and the platform's tools expose students to practices relevant to research and industry; professional proficiency is not claimed.Source: B. Platform · p. 3
- Learners progress from shell choices to bounded motion goals and reward changes, testing new motions in simulation before any supervised hardware attempt.Motors and safety settings remain fixed; the training code is scaffolded, and the activities' difficulty remains to be evaluated.Source: B. Platform · p. 3
C. Track Organization
- Four coordinated tracks connect disciplinary ideas to practical decisions: ME links geometry and balance to predictions, Build develops assembly and fault-diagnosis practice, EE emphasizes safe testing, and CS+AI examines models and rewards.The tracks use simulation, learning, and physical comparison to examine how models and rewards shape behavior.Source: C. Track Organization · p. 3
- On-robot control requires an electrical power-on gate, and walking requires a standing test, while simulation can begin independently to support early prediction and interpretation.Fig. 2 depicts students collaborating during assembly, soldering, and on-robot testing.Source: C. Track Organization · p. 3, Figure 2 · p. 3
D. CS+AI Activities and Checkpoints
The CS+AI plan combines simulation, policy learning, and supervised physical trials with individual explanations. Checkpoints distinguish robot performance from conceptual understanding and include a fallback for unavailable hardware.
D. CS+AI Activities and Checkpoints
- Students use simulation for repeated trials, learn from demonstrations and rewards, use PPO to update a policy, and vary simulated conditions to prepare for physical differences.Students test a supplied policy before their own, and all physical trials require mentor supervision.Source: D. CS+AI Activities and Checkpoints · p. 3
- Each learner predicts the effect of a reward change, then the pair trains and compares behavior with the original while each student explains whether higher rewards mean better behavior.The instructor uses those explanations to provide support or an extension.Source: D. CS+AI Activities and Checkpoints · p. 3
- Hardware targets include standing and approximately five seconds of walking with the supplied policy, followed by an attempt with a student-trained policy.Logs should distinguish supported stepping from unsupported walking; neither demonstrates individual conceptual mastery.Source: D. CS+AI Activities and Checkpoints · p. 3
- When hardware is unavailable, simulation comparisons and measurement-log analysis preserve conceptual work but do not replace the safe-startup outcome; three-hour sessions alternate primers, supervised labs, and debriefs.Long training runs occur between sessions, short physical trials and spare parts accommodate fall damage, and workload and reliability remain unevaluated.Source: D. CS+AI Activities and Checkpoints · p. 3
A. Protocol and Analysis
Five experts reviewed the initial curriculum through structured interviews, ratings, and open-ended prompts, with ratings summarized descriptively and responses coded into themes. The materials distinguish that initial review from the proposed revisions, which experts did not evaluate.
A. Protocol and Analysis
- Five experts from robotics research and engineering, secondary STEM teaching, and curriculum design reviewed the initial four-track structure, eight-session CS+AI plan, safety checkpoints, and staged deliverables.The panel's curriculum-relevant backgrounds are identified in Table II.Source: A. Protocol and Analysis · pp. 3–4, Table II · p. 4
- Ratings and qualitative findings concern the initial curriculum, not the revised session plan or framework as a separate instrument; the three tensions are an author synthesis.Table III presents proposed revisions that have not been re-evaluated.Source: A. Protocol and Analysis · p. 3, Table III · p. 5
- Fig. 3 depicts conceptual progression across ME, Build, EE, and CS+AI, linking robot performance with individual understanding; safety checks constrain hardware work while simulation can begin independently.The caption notes that this is not a synchronized timetable.Source: Figure 3 · p. 4
- Table I identifies a revised eight-session CS+AI plan as proposed, with provided code supporting core tasks and independent implementation and student-trained walking as extensions.The plan is explicitly labeled not re-evaluated.Source: Table I · p. 4
- The adapted rubric rated eight curriculum dimensions on a five-point scale and used open-ended prompts; ratings were summarized descriptively and responses iteratively coded into cross-case themes.One expert's missing overall-readiness score was not imputed, and a duplicated interview record was counted once.Source: A. Protocol and Analysis · p. 4
- Fig. 4 displays expert ratings across eight dimensions for five experts, using bars for means, whiskers for sample standard deviations, and circles for individual five-point ratings.Source: Figure 4 · p. 5
B. Quantitative Findings
Experts rated integration and teamwork favorably, while their qualitative feedback identified six themes about coherence, cognitive load, resources, participation, reasoning, and relevance.
B. Quantitative Findings
- Both integration of AI/robotics content and teamwork averaged 4.60 (SD = 0.89); communication was comparatively weaker at 3.40 (SD = 0.55).Among four experts providing an overall score, implementation readiness averaged 4.00 (SD = 0.82, range 3–5).Source: B. Quantitative Findings · p. 5
C. Qualitative Findings
- All five experts valued connecting assembly, electronics, simulation, and deployment through one robot, while preferring a smaller conceptual core over equal disciplinary depth.Source: C. Qualitative Findings · p. 5
- Four experts questioned terminology, prerequisites, long sessions, and setup work that could delay visible success; a motivation rating of 1/5 cautions against reading the favorable mean as consensus.Source: C. Qualitative Findings · p. 5
- All five experts identified implementation conditions such as reliability, setup, expertise, supervision, or cost.Suggestions included validated instructions, configured software, and backup activities.Source: C. Qualitative Findings · p. 5
- Experts supported pair work but warned that differences in experience could concentrate technical work in one student; they suggested complementary pairs and rotating roles.One expert could not confidently judge teamwork because individual responsibility was underspecified.Source: C. Qualitative Findings · p. 5
- Experts distinguished passing a robot milestone from understanding it, calling for explanations of failures and design decisions and interpretation of logs and measurements.Source: C. Qualitative Findings · p. 5
- Three experts recommended linking concepts to observable behavior and future study opportunities, with demonstrations and frequent achievements to sustain engagement.Source: C. Qualitative Findings · p. 5
D. Design Tensions and Framework Refinements
The authors synthesized expert feedback into three design tensions and revisions. The proposed individual assessment records each learner’s reasoning and practical evidence separately from pair robot performance.
D. Design Tensions and Framework Refinements
- The authors frame three tensions synthesized from six expert-feedback themes; Table III documents design changes, not learning gains.Source: D. Design Tensions and Framework… · p. 5
- At each checkpoint, each student submits a separate prediction and explanation, completes an observed practical task, and answers a brief follow-up without partner assistance.The instructor records whether performance was independently demonstrated, prompted, or not yet demonstrated, with supporting evidence.Source: D. Design Tensions and Framework… · p. 6
- Individual judgments are recorded separately from pair robot performance, so partners need not receive the same assessment.The procedure is proposed and remains to be evaluated.Source: D. Design Tensions and Framework… · p. 6
VI. DISCUSSION AND CONCLUSION
The discussion limits the evidence to formative expert feedback and identifies resource and transfer conditions that future classroom studies must examine.
VI. DISCUSSION AND CONCLUSION
- Five purposively selected experts provide formative feedback, not representative agreement; ratings establish neither formal content validity nor learning effectiveness.The revised framework and assessment criteria remain unevaluated.Source: §VI · p. 6
- Transfer to other platforms requires adapted prerequisites, checkpoints, and assessments, while mentoring, training compute, and repair capacity remain resource assumptions.Classroom studies must measure workload and success rates.Source: §VI · p. 6
- Comparisons with simulation-only RL should account for prior experience, contact time, and instructor support; any learning advantage remains untested.Suggested shared tasks assess reward design, training versus evaluation, unfamiliar failures, completion time, assistance, and student explanations.Source: §VI · p. 6