GYOZA
Measure RoboCasa reset/step and held-out success before proceeding to RL.
Yugo Ito
High school student working on robot learning, embodied AI, and simulation.
I build small-scale research projects around manipulation, synthetic data, and world representations.
Current research
Code-as-World
Recheck the submitted v3 result before claiming model capability.
Synthetic Data in Robot Learning
Test the perception-bottleneck reproduction before fixing relative baselines.
Latest notes
Research record
GYOZA
A robot-learning project for tabletop food manipulation, orchestration, and verification.
- Role
- Design, experiment, and verification of VLM×ACT autonomous robot-arm systems, centered on sim-to-real and human-video transfer.
- Evidence
- Ran manipulation experiments on the physical SO-101 arm (median RTT 658ms)
- Passed the MisoBowlToTray smoke test in the RoboCasa environment
- Recorded a no-go decision on Web-video pseudo-actions in GYOZA-Scaling
- Next
- Measure RoboCasa reset/step and held-out success before proceeding to RL.
Code-as-World
A world-model direction that treats environments as executable scene descriptions.
- Role
- Designed the conditional world-model gate evaluation and a safe deterministic evaluator.
- Evidence
- Implemented the Stage 0A white-ball world-code benchmark
- Verified cwd via HF Jobs / Pi integration
- Confirmed Linux EGL/OSMesa as the rendering acceptance path
- Next
- Recheck the submitted v3 result before claiming model capability.
Synthetic Data in Robot Learning
Experiments on when synthetic data helps behavior—and when perception becomes the bottleneck.
- Role
- Designed the synthetic-data pipeline and experiments separating behavior from perception.
- Evidence
- Recorded the observation that perception, not behavior, becomes the bottleneck in synthetic-data learning
- Experiment logs to be published in Notes
- Next
- Test the perception-bottleneck reproduction before fixing relative baselines.
Previous projects
Shiin IME
An iOS custom keyboard that types Japanese from consonants only, with on-device Core ML inference.
- Role
- Designed and implemented the consonant-to-reading model, Core ML export, and iOS keyboard extension.
- Evidence
- Released the implementation under the MIT License
- The public repository documents a roughly 1.3M-parameter Transformer encoder + GRU decoder and the on-device Core ML design
TeenEmo
A training pipeline for Japanese emotional-intelligence language models: data preparation, SFT/DPO, and evaluation.
- Role
- Designed and implemented the synthetic-data, SFT/DPO, Japanese EQ-Bench3 translation, and evaluation pipeline.
- Evidence
- The public repository records the path from data generation through training and evaluation
- Released a DPO model and a Japanese EQ-Bench3 dataset on Hugging Face
Lumi
A hackathon prototype of an AI voice companion for older adults, using dementia care as its design context.
- Role
- Implemented the prototype's QLoRA training, voice interaction pipeline, conversation memory, and scam-deflection layer.
- Evidence
- The public repository documents QLoRA training on AMD MI300X and the voice-interaction pipeline
- Released a model and Gradio Space on Hugging Face
quiz-ai
A Japanese competitive buzz-quiz LLM system that plays under the same conditions as a human.
- Role
- Designed and implemented the question corpus, answering and buzz-timing models, training/evaluation path, and ZeroGPU demo.
- Evidence
- The public repository documents the two-model design for character-by-character buzz decisions and the training/evaluation path
- Prepared a public Gradio Hugging Face Space
Links
Public work and profile links.