I'm Wang Haolin, an AI major at Sichuan University (SCU) in Chengdu. I build LLM agents and the harnesses that drive them, and push embodied AI from simulation onto real legged robots and robot arms โ with a side track in RL for high-frequency markets.
- ๐ Artificial Intelligence @ Sichuan University ยท Chengdu, China
- ๐ค LLM agents & harnesses โ multi-agent pipelines, MCP tooling, skill systems, context engineering
- ๐ฆฟ Embodied AI / VLA โ RL locomotion (PPO, imitation, AMP), ฯ0.5 VLA deployment on edge devices
- ๐ RL for markets โ market making & microstructure-based direction prediction at tick level
- ๐ง Methods I use daily: PPO ยท imitation learning ยท domain randomization ยท teacher-student distillation ยท sim2real
- ๐ซ Reach me: whl@scu.edu.cn
LLM-driven rounds: trend analysis โ spec evaluation โ code generation with test-driven self-healing โ deployment โ engagement, orchestrated end-to-end by a scheduled pipeline.
|
Transparent proxy that compresses MCP tool definitions & responses โ saves up to 80% of context-window tokens for AI coding agents.
|
One command to create, test, publish and install agent skills, hooks and sub-agent configs โ the npm for AI agent capabilities.
|
Describe a robot in plain English, get a simulation-ready MuJoCo scene with physics, assets and RL training scripts in seconds.
|
- ฯ0.5 VLA real-robot deployment โ LoRA-finetuned ฯ0.5 (PaliGemma 2B + action expert) running closed-loop on NVIDIA Jetson Thor ร AgileX Piper arm: natural-language pick-and-place with 10-step action chunks, full perception โ inference โ control loop
- DreamWaQ for wheel-legged robots โ adapted DreamWaQ (PPO + ฮฒ-VAE implicit terrain imagination) to 4ร4 wheel-legged platforms: up to 1 m step climbing, Isaac Gym โ ONNX โ MuJoCo sim2sim โ C++ deployment on real robots
- Backflip on a 60 kg quadruped โ DeepMimic-style imitation on ZJ-T30-V2: trajectory-optimization reference + PPO tracking, phase-driven control, motor torque-speed envelope
- AMP & CAMP multi-gait control โ adversarial motion priors for natural trot, rough-terrain curriculum and fall-recovery; skill-conditioned CAMP learns walk / trot / pronk / bound in a single policy
- Extreme Parkour โ end-to-end depth-vision parkour on Unitree Go2: teacher-student privileged RL, depth-latent + heading distillation, ROS 2 deployment
- HIMLoco for Go2W โ human-imitation locomotion (ICLR 2024) re-implemented on the wheel-legged Go2W: Isaac Gym training + MuJoCo sim2sim
- AI by Hand (Excel) โ deep learning rebuilt from scratch in spreadsheets: backprop, RNN/LSTM/xLSTM, ResNet, full-stack Transformer, Mamba, all the way to AlphaFold โ the best way to really understand the math
- RLMM โ reinforcement-learning market making on Binance ETH-USDT perps: tick-level
hftbacktestGym environment, RecurrentPPO + LSTM with a 4-phase curriculum over ~900 days of L2 orderbook data; plus a 66-dim microstructure direction model (84.5% validation accuracy, profitable on 93/100 out-of-sample days)
- ๐ค Agent harness engineering โ long-running loops, context budgeting, skill ecosystems
- ๐ง VLA on the edge โ shrinking ฯ0-class models onto Jetson-class devices with usable control rates
- ๐ฆฟ Wheel-legged sim-to-real โ taking Go2W RL policies from Isaac Gym onto the real robot
- ๐๏ธ Agile skills โ imitation & contrastive learning for parkour-class maneuvers
โก "็ฅ่กๅไธ โ knowledge and action as one." โก