RoboStudy

Inspiration

Robotics study material is split between a hands-on Playbook, lab guides, PDFs and slides. A useful study companion should show where an answer comes from and help a student practice what they have not mastered, rather than produce a convincing answer with no evidence.

What it does

RoboStudy is a Track D prototype built on public robotics course material. The public draft supports source search over 95 cited content units, exact excerpt viewers, local PDF and PowerPoint import, original PDF-page and embedded slide-image viewers, 12 source-validated authored questions, balanced five-question mock exams, cited feedback and a device-local learner dashboard. It declines unsupported searches instead of inventing course content.

How it works

A Vite and JavaScript frontend uses PDF.js for PDF text and exact-page rendering, and JSZip for slide text and embedded images. Source units keep section, page, slide or timestamp provenance. Lexical retrieval selects evidence; optional generated answers must quote actual source text and reference valid unit IDs. Practice prioritizes weaker topics and unseen questions. A Bayesian knowledge-tracing estimate updates after answers, with lower weight for transparent keyword-graded short answers. Conversation engagement is recorded separately, not counted as proof of mastery.

An optional Python Gemini gateway keeps the key server-side and billing disabled. Local tests passed PDF-diagram understanding, a real two-slide PPTX, and a small spoken-video fixture with timestamped content. Generated questions require a separate verification pass before acceptance. Concept-map output requires exact quotes, unique nodes, valid edges and an acyclic graph.

Current limits and challenges

This is a public draft, not a finished Track D system. The hosted static app has live AI disconnected. Free-tier model requests were intermittent and returned timeout/503 errors; the latest question verifier and live prerequisite-map test did not pass. Large lecture videos currently use visual sampling rather than verified full-lecture audio. Local PowerPoint images are preserved, but original slide layouts are not rendered. Current upload topic tags use transparent keyword rules.

Evaluation

A seven-case deterministic retrieval baseline includes known-source and off-material queries. One actual /clock tutor answer completed built-in DeepEval faithfulness, answer-relevancy, context-precision and recall metrics, with per-metric judge provenance. That is a smoke test, not broad model accuracy. A full multi-case semantic test set remains unfinished. Three scripted learner profiles run across multiple sessions without repeating question IDs; changing mastery estimates does not prove human learning gains.

What I learned and next steps

Evidence validation and graceful fallback matter when a model service is unreliable. Next steps are a reliable free-tier hosted AI path, verified long-lecture ingestion, completed generation-and-verification tests, and a wider semantic evaluation set before stronger performance claims.

Credits and AI assistance

Course sources: the public Robotics Playbook and Isaac Sim/ROS 2 Lab 3. Lab 3 is co-authored by Prajval Arora and Sai Srinivas Tatwik Meesala for Prof. Sangram Redkar. No private course originals are bundled. Code, UI, tests and documentation were AI-assisted by Instinct. Gemini was used in optional local inference/evaluation tests. No model fine-tuning was performed.

Built With

• deepeval
• gemini
• instinct
• javascript
• jszip
• pdfjs
• python
• vercel
• vite