A desktop robot that recognizes objects, responds to commands and uses human feedback for cat-like trial and correction.
My role Team member · Visual perception, multimodal interaction and system integration
The robot was built as a team project. My work covers perception, interaction and integration, rather than sole ownership of mechanics, electronics and drivers.
YOLOv8
OpenCV
Vosk
MediaPipe
Contributions, results & full validation scope
My contributions
Trained and deployed a YOLOv8 detector for three tabletop object classes and organized image capture and target localization.
Used OpenCV for camera calibration, distortion correction and homography mapping to the work plane.
Integrated Vosk Chinese speech, keyboard input and MediaPipe hand detection into trial, correction and reward interactions.
Connected perception, coordinate conversion and interaction decisions to the team's robot execution interface for picking and placement.
Recorded results
The historical training CSV records a best validation mAP@0.5 of 86.21%, a detection metric.
Stored camera reprojection RMS is 0.704 px; the eight-point homography fit has an RMS residual of 7.145 mm.
The physical project was completed and handed in; the public archive retains application code, calibration files and training records.
Validation scope & limitations
The robot was built as a team project. My work covers perception, interaction and integration, rather than sole ownership of mechanics, electronics and drivers.
Detection and calibration metrics are not grasp success rates; no fixed-protocol physical repeatability study is available.
Archive checks were offline only, with no new training, inference or robot motion. Weights, datasets and vendor DLLs are excluded.
Multicamera vision-language-action policy adaptation, training deployment, local closed-loop simulation and failure analysis.
My role Individual competition project · Open-model adaptation, experiment organization and policy deployment
StarVLA, QwenOFT, XPolicyLab and RoboDojo provide the upstream foundations. My work is adaptation, configuration, integration, runtime fixes and analysis.
StarVLA
QwenOFT
WebSocket
Isaac Sim
Contributions, results & full validation scope
My contributions
Adapted head and wrist camera inputs, language instructions, and 14-dimensional robot state and action interfaces.
Configured action-head fine-tuning and later multitask training, managing checkpoints, normalization settings and frozen versions.
Connected a WebSocket policy service to multiple simulation environments and organized recovery checks and evaluation guardrails.
Organized two full evaluations, candidate-checkpoint screening and task-level failure analysis with traceable results and video indices.
Recorded results
The stable baseline achieved 33/600 and the 20k fine-tune 28/600 across 24 tasks each: 1,200 episodes in total.
Candidate C checkpoints at 60k, 80k and 100k received 45 screening episodes in total, each scoring 0/15; no full 600-episode reevaluation was completed.
Diagnostic material covers 29 representative episodes and an index of 87 three-camera videos.
Validation scope & limitations
StarVLA, QwenOFT, XPolicyLab and RoboDojo provide the upstream foundations. My work is adaptation, configuration, integration, runtime fixes and analysis.
These are local simulation records without a confirmed official score. The 20k aggregate observation is below baseline and does not establish improvement.
The 45-episode screening is not directly rankable against full 600-episode evaluations; no performance advantage is established for the 100k candidate.
Archiving did not rerun GPU training or full simulation. Raw 20k episodes, complete ACT source and implementations for some historical tests remain incomplete.
Five kernel filters, historical linear fingerprint experiments and reading notes on Transformers for indoor localization.
My role Course reproduction and implementation organization · Interfaces, numerical safeguards and teaching notes
Measured data were supplied and preprocessed by the teacher and are excluded. Full course experiments and hardware collection were not repeated during archiving.
MATLAB
KRLS
Wi-Fi RSSI
Attention
Contributions, results & full validation scope
My contributions
Organized common interfaces for five kernel filters and coordinate experiments, keeping the 70/20 and 40/16 data protocols separate.
Added a portable CSV entry and micro-synthetic checks, and organized Attention and AaT reading notes.
Recorded results
Micro-synthetic MATLAB checks passed for the five kernel filters and the Wi-Fi component.
A shared entry organizes algorithms, data contracts and Transformer reading material.
Validation scope & limitations
Measured data were supplied and preprocessed by the teacher and are excluded. Full course experiments and hardware collection were not repeated during archiving.
The algorithms are educational reproductions; the Transformer component is explanatory. Historical curves do not establish cross-site accuracy or algorithm rankings.
Historical RTL, state control and limited behavior tests for understanding digital-system versions and interfaces.
My role Course exercises and group material · No sole-authorship claim for all RTL
Some tests bypass keypad scanning and initialize internal registers. The main chess-clock version needs a temporary compatibility copy; this is not full-system acceptance.
Verilog
Icarus
FSM
XDC
Contributions, results & full validation scope
My contributions
Organized multiple source and constraint versions for two-player timing, keypads, displays and password control.
Added separate-version compilation, short controller simulations and defect reproduction while preserving original RTL.
Recorded results
Version-level Icarus compilation checks and limited timing and controller behavior checks were completed.
The binary password version covers 128 input combinations; original compile failures and unknown-state startup defects remain explicit.
Validation scope & limitations
Some tests bypass keypad scanning and initialize internal registers. The main chess-clock version needs a temporary compatibility copy; this is not full-system acceptance.
No complete keypad/display chain, Vivado timing closure or board acceptance is established; original pin constraints still need review.
Coursework from state feedback, observers and digital-control models to a traffic-light STATE3 ladder module.
My role Course experiments; STATE3 design and debugging within the PLC team
Short simulations do not establish full settling behavior, ripple-free response or robustness; historical coefficient issues remain.
MATLAB
Simulink
PLC
Micro850
Contributions, results & full validation scope
My contributions
Organized state-feedback, observer, digital-PID and deadbeat-control scripts and models.
Retained my east–west-through STATE3 module as eight ladder rungs with I/O checks; the complete state machine belongs to the team.
Recorded results
MATLAB numerical checks passed and three original Simulink models each ran to 0.6 seconds.
STATE3 image, transcription and timer-parameter archive-consistency checks passed.
Validation scope & limitations
Short simulations do not establish full settling behavior, ripple-free response or robustness; historical coefficient issues remain.
No native PLC project or new compilation, download or device test is available. STATE3 does not establish complete interlocks, yellow lights or safe stopping.
Historical circuit construction, parameters and waveform observations for communication modulation and literature-based gate driving.
My role Analog-communication team leader; simulation and reporting for the isolated-drive project
Native projects are excluded and no new Multisim or hardware run was performed; screenshots do not establish SSB suppression or insulation performance.
Multisim
MC1496
Filters
Gate Drive
Contributions, results & full validation scope
My contributions
Participated in the team's Wien-bridge, MC1496 modulation and filter design; individual module assignments are unconfirmed.
Built an approximate literature-based Multisim circuit and handled simulation and reporting for the isolated-drive coursework.
Recorded results
Original circuit screenshots, parameter tables and historical waveform records are public.
Parameter, image and file-consistency checks passed without newly solving the circuits.
Validation scope & limitations
Native projects are excluded and no new Multisim or hardware run was performed; screenshots do not establish SSB suppression or insulation performance.
The drive screenshots do not confirm a SiC model and specify only a nominal 10 Hz pulse; high-frequency performance and synchronization delays are unsupported.