Jeff Liu

Jeff Liu

Also known as Yanshifu

Experiment Reports

Jeff Liu, also known as Yanshifu, is a co-founder of BAP Lab and Yanshi Robotics. His work focuses on reproducing cutting-edge research, translating it into practical engineering applications, and supporting its wider adoption. He also develops robots that address real-world problems.

Experiment Reports

Experiments in robotics, machine learning and intelligent systems. Methods, observations and limitations, preserved alongside the original reports.

2026

SO-101 snack pick-and-place with ACT

· PDF report · English

An ACT study connecting 90 dual-camera leader–follower demonstrations, network architecture, training logs and two autonomous physical pick-and-place trials.

Reinforcement & robot learningRobotics

29-DoF G1 locomotion and Sim2Sim transfer

· Interactive report · Chinese

An asymmetric actor–critic with deployment-format observations is trained in Isaac Lab and transferred to MuJoCo, documenting cross-engine control debugging.

Reinforcement & robot learning

Failure cases: training metrics versus displacement

· Interactive report · Chinese

An analysis of G1 policies with favorable training metrics but inadequate commanded motion, using measured displacement to examine capability claims.

Reinforcement & robot learning

G1 speed ladder: walking to running

· Interactive report · Chinese

Expanding the forward-command range on the stock G1 to examine velocity tracking, aerial phases and stability.

Reinforcement & robot learning

Flat-ground locomotion with the stock G1

· Interactive report · Chinese

Velocity tracking and entropy-coefficient ablations on the stock Isaac Lab G1, with multi-seed training and independent simulation evaluation.

Reinforcement & robot learning

Ant locomotion and ablation experiments

· Interactive report · Chinese

Learning locomotion on the eight-joint Ant, comparing energy penalties and entropy regularization with multi-seed experiments.

Reinforcement & robot learning

Cartpole balance with PPO

· Interactive report · Chinese

PPO training for the Isaac Lab cartpole task, connecting observations, actions, rewards, networks and training curves.

Reinforcement & robot learning

VisInject: image prompt injection in VLMs (v1.1)

· PDF report · English

A three-stage study combining universal adversarial-image optimization with AnyAttack fusion, evaluating 6,615 clean/adversarial response pairs and separating output disruption from target-content injection.

Computer visionDeep learning

Weather forecasting: CNN baseline training

· Training report · Chinese / English

A six-variable regression CNN trained on single-frame weather fields, documenting architecture, learning curves, validation errors and overfitting.

Weather forecastingDeep learning

2025

Meta-learning on Mini-ImageNet

· PDF report · English

Mini-ImageNet experiments compare MAML with a fine-tuning baseline and examine class count, shots and inner-loop update steps.

Deep learning

From simplified environments to Atari Breakout

· PDF report · English

A reinforcement-learning study progressing from Q-learning in a simplified brick game to DQN in Atari Breakout, with implementation, debugging and training observations.

Reinforcement & robot learningDeep learning

An AI agent prototype for sepsis analysis

· PDF report · English

A language-model agent coordinates constrained analysis tools, sepsis data processing and result explanations, with a discussion of layered agent design.

Agents & systems

Color calibration with machine learning

· PDF report · English

A calibration-card pipeline maps photographed colors to reference colors, documenting data preparation, feature construction, model comparisons and deployment.

Computer visionDeep learning

2024

Centroid-based robot line following

· PDF report · English

OpenCV locates a colored line centroid for proportional control, with search and obstacle-avoidance states for interrupted tracking.

RoboticsComputer vision

Maze escape with left-wall following

· PDF report · English

A sensor-based maze navigation study using laser scans, left-wall following and a state machine, with obstacle handling and simulation tests.

Robotics