FEATURED NOTE / IN PROGRESS
A teacher is only useful when the dataset remembers why it worked.
From stable demonstrations to a LeRobot dataset: the first loop I want to make boring, repeatable, and measurable.
IMITATION LEARNINGLEROBOT
08 MIN READMANIAC1UM / TECH NOTES
Building machinesthat learn fromthe world.
A working notebook on embodied AI, imitation learning, MuJoCo, LeRobot, and the small engineering decisions that make Sim2Real experiments move.
OPEN NOTEBOOK / 2026
01 / NOTEBOOK
Selected field notes
Short, honest notes from the workbench — what worked, what broke, and what I am trying next.
02
Sim2Real starts with an evaluation loop, not a lucky checkpoint.
A field guide to separating scene setup, action schema, contact checks, and the policy itself.
03
The safety boundary belongs inside the embodiment stack.
Notes on constrained scenarios, failure modes, and why a moving system needs more than a good average score.
02 / BUILD LOG
Projects in motion
A small set of open experiments — each one is a place to test an idea in public.
01ACTIVE
MuJoCo Dexterous VLA
A reusable dexterous manipulation platform that connects Teacher demonstrations, LeRobot datasets, VLA fine-tuning, and policy evaluation.
02COMPLETE
OpenClaw Market Analyser
A market analysis tool for data collection, signal processing, and an interactive dashboard built for research and monitoring.
03ACTIVE
Paper Repro
An agent workflow for parsing papers, extracting methods, rebuilding experiments, and validating reproducibility in stages.
03 / OPEN QUESTIONS
Research trail
Questions I keep returning to as the projects get more physical.
DATA BEFORE DEMOS
What makes a demonstration worth keeping?
Looking at the small choices — contacts, resets, action dimensions, and failure labels — that decide whether a dataset teaches or merely records.
POLICY IN THE LOOP
Can evaluation be designed before training?
Treating gates, viewers, headless runs, and regression baselines as part of the policy rather than a report written afterwards.
EMBODIED SECURITY
Where does a physical system fail safely?
Exploring constrained environments where a model must be useful, bounded, and explainable when the average case is not enough.
04 / THE PERSON BEHIND THE LOG
About the notebook
I am Maniac1um, a cybersecurity student at Hangzhou Dianzi University learning by building. This site is where I keep the experiments, notes, and unfinished questions that sit between a paper and a working system.
Send a noteEDUCATIONB.S. / School of Cyberspace Security2024 — PRESENT
FOCUSEmbodied AI · Imitation learning · Sim2Real↗
TOOLKIT
PythonPyTorchMuJoCoLeRobotLinuxSecurity
06