Building a complete smartphone-teleoperation + imitation-learning stack on a low-cost 6-DoF arm — from first principles, every component understood, every artifact reproducible. Target: ICRA 2027 (backup) / ICRA 2028 (primary).
Graduate student in Vietnam, self-directing a research track in robotics and machine learning with Tuan Dang (University of Arkansas) as advisor.
I study every topic from first principles — no black-box libraries until I understand what's inside. Every topic must leave a runnable artifact in the repo before I move on.
Why I can do this
Live working-notes dashboard with CI auto-generation (this page) already running
5 deep technical notes completed across kinematics, servo protocols, signal filtering, IK, VIO drift
Weekly accountability structure with a university advisor from day one
Caught a prior-work overlap (Phone2Act, May 2026) early and re-planned — preventing wasted experiments
What I am building toward
A peer-reviewed paper at ICRA 2027 (early attempt if results ripen) or ICRA 2028 as primary target, on a controlled comparison of smartphone vs. conventional teleoperation interfaces × data quality × policy performance on a low-cost arm.
Open-source commitment
The full teleoperation stack — code, reproducible README, demo video, and a technical blog post — will be released open-source by end of 2026, regardless of publication outcome. This is a forcing function, not a promise contingent on success.
Active projects
SO-101 · Smartphone Teleoperation → ICRA
active
WebXR streams phone pose at 60 Hz → 50 Hz Python server (clutch delta-pose, One-Euro filter, DLS-IK) → 6-DoF SO-101 arm. Demonstrations recorded as LeRobot datasets → ACT / SmolVLA imitation learning. Target: ICRA 2027/2028.
This project runs on a student budget. Hardware is the main bottleneck — compute is covered by Kaggle free tier. If you see value in open, reproducible, from-scratch robotics research from Southeast Asia, here is how you can help.
Hardware budget — what is needed
The SO-101 arm kit, servos, control board, PSU, and 3D-printed parts total roughly $500–700 USD (one-time cost). All code and data produced will be open-source. Compute beyond Kaggle's free tier (if needed for ablations) adds ~$100–200.
If you sponsor hardware, you get named in the repo README, the open-source release post, and (if the paper is accepted) the acknowledgments section.
Tier 1
Research mentor
Researcher or engineer who wants to co-steer the project — weekly 30-min calls, paper co-authorship possible. No money required.
Tier 2
Lab / grant sponsor
A university lab or grant that covers hardware costs (~$700). In return: acknowledgment, open-source data, and a co-authorship conversation.
Tier 3
Hardware / company
A robotics company (arm kits, servo suppliers, compute) that donates hardware or credit. Named in all releases and the paper. No equity, no exclusivity.
Tier 4
Community supporter
Anyone who believes in open robotics research — Ko-fi / GitHub Sponsors. Even $10 moves the hardware timeline forward. All supporters named in the release post.
→ Reach out at sunshineforwho@gmail.com — happy to share a detailed project brief, budget breakdown, or join a call.
Skills — each one tied to evidence
A skill is only claimed when the milestone that proves it is done. Gray = scheduled, with the phase that will produce the evidence. Full evidence on the project page →
Mathematics & robotics
3D rotations — SO(3), quaternions, Lie exp/logplanned
Forward / inverse kinematicsplanned
Differential IK — damped least squaresplanned
Signal filtering — One-Europlanned
PID & trajectory trackingMar – Jun 2027
Machine learning
PyTorchNov 2026 – Feb 2027
Neural nets from scratch (Karpathy Z2H)Nov 2026 – Feb 2027