Ego-centric Video World Model (WanFunHands): On top of the Wan2.2-Fun-5B base, independently designed a hand-conditioning injection architecture (skeleton channel-concat + HPP token injection + camera LoRA, all zero-initialized to preserve base priors). Processed 3,600 hours of ego-centric data into trainable multimodal condition latents and delivered usable 480p and 720p model weights with full inference/training scripts and demos.
NitroGen Semantic Data Engine: Independently built a production-grade distributed annotation pipeline (~15k lines) for NVIDIA NitroGen game data, supporting ~10-machine sharded parallelism, checkpoint resume, and exponential-backoff retries. Produced ~1.3M training samples (~3,500 hours) of high-quality image–action aligned data via three-channel VLM semantic reasoning.
P2P → NitroGen Format Conversion: Built a full conversion pipeline mapping keyboard/mouse data to gamepad format, with a routing system covering 7 game categories and 94 combinations. Implemented mouse→right-stick PD+EMA smoothing and W→trigger throttle ramps, converting all 108,851 clips (~10k hours) with a three-tier validation system.
Embodied Intelligence InternACE RoboticsNov 2025 – Jan 2026 ▶
Data Pipeline: Developed a high-efficiency pipeline converting heterogeneous spatial intelligence datasets (3D Grounding, 6DoF) into Qwen SFT formats.
Model Training & Evaluation: Executed full-parameter fine-tuning on Qwen3-VL and established a comprehensive benchmark evaluation system for embodied tasks.
Unitree G1 Guidance System: Designed and optimized robotic motion libraries for guided navigation, integrating custom action sequences with real-time kinematic control.
LLM-Driven Data Production Pipeline: Explored an LLM-driven data production pipeline using NVIDIA Isaac Sim API to automatically generate USD files from scene configs for simulation training environment construction.