ROBOTICS & EMBODIED INTELLIGENCE
I'm Zheng Wu, an undergraduate exploring robot learning, physical control, and visual perception. These are my experiments at the intersection of algorithms and the real world.
Explore selected work ↓Beijing Jiaotong University / RIT
Undergraduate · Expected graduation 2027
A SELECTION OF HANDS-ON PROJECTS
01ROBOT LEARNING / SIMULATION
Quadruped reinforcement learning · Isaac Lab
Exploring reinforcement learning for quadruped locomotion in Isaac Lab, with a recorded demonstration from the simulation environment.
A simulation-based robot learning project. This recording presents the current demonstration; training settings and quantitative evaluation will accompany a future project update.
02ROBOT LEARNING / MANIPULATION
SO101 · Imitation learning & VLA
Real-world manipulation experiments with the open-source SO101 arm, connecting demonstration collection, policy training, and deployment on physical hardware.
FROM EXPERIMENT TO RESEARCH INTEREST
I began this project by training an ACT policy on approximately 60 teleoperated demonstrations for a grasping task with the SO-101 robotic arm. The policy worked well near the demonstrated object positions, but became less reliable as I moved the object away. This raised a question: how much of the learned grasping behavior would transfer beyond the spatial distribution represented in the demonstrations?
To investigate, I systematically varied the object position using the same checkpoint, while aiming to keep the robot configuration, camera setup and object unchanged. In these small-sample tests, grasping was relatively stable within 0–4 cm. At left 7 cm, both runs eventually succeeded, but required four and six attempts; at left 8 cm, the single recorded run failed. This suggested that training-data coverage could constrain generalization—and that final success alone could hide declining efficiency.
I then collected demonstrations from underrepresented regions and fine-tuned the original policy for 10,000 steps using only the new data, because incompatible observation schemas prevented merging the datasets. The resulting policy succeeded at several newly covered positions, but also failed at some nearby positions. This raised the possibility of a shift in the policy’s spatial preferences or forgetting. Because the two evaluations used different position sets and few trials, I could not yet establish either a workspace expansion or a loss of previous capabilities.
This shifted my perspective. The challenge was no longer only how to adapt to new situations, but also how to preserve existing capabilities while learning from new experience. Thinking about robots encountering changing environments, objects and tasks led me to become interested in Continual Robot Learning: how robots can balance adaptation and retention, and accumulate capabilities over time. A next step is to compare new-only fine-tuning with mixed old-and-new data on the same evaluation grid.
PROJECT DEMONSTRATIONS
Set up the robotic arm, collected demonstrations, trained ACT imitation learning policies, and experimented with VLA models using visual observations and language instructions.
FROM OBSERVATION TO THE NEXT TEST
SO101 · ACT block grasping
Exploratory experiments / Day 1–2
DAY 1 / DIAGNOSE
EVIDENCE / LEFT 7 CM
The two runs took 4 and 6 attempts. At the origin, all three runs succeeded on the first attempt.
| Position | Runs | First success | Final success |
|---|---|---|---|
| Origin | 3 | 3 | 3 |
| Left 2 | 3 | 2 | 2 |
| Right 2 | 2 | 2 | 2 |
| Left 4 | 3 | 2 | 3 |
| Right 4 | 2 | 2 | 2 |
| Left 7 | 2 | 0 | 2 |
| Right 7 | 3 | 2 | 3 |
| Left 8 | 1 | 0 | 0 |
Left 4 cm included one run requiring three attempts; right 7 cm included one requiring two. The notes mention right 9 cm failures without a tabulated count, so that position is excluded here.
DAY 2 / INTERVENE
EVIDENCE / AFTER FINE-TUNING
Eventual successes / runs. Left 8 cm failed in all three runs; left 9 + up 3 cm succeeded in all three.
| Position | Runs | Final success |
|---|---|---|
| Left 7 | 3 | 2 |
| Left 8 | 3 | 0 |
| Left 8, up 2 | 1 | 1 |
| Left 9, up 3 | 3 | 3 |
| Right 5, up 4 | 3 | 3 |
| Left 3 | 2 | 0 |
| Right 3 | 3 | 1 |
| Right 7 | 3 | 3 |
| Right 7, up 3 | 3 | 3 |
| Right 7, down 1 | 3 | 3 |
| Up 2 | 3 | 0 |
The right 3 cm success required two attempts. Day 1 already achieved 3/3 eventual successes at right 7 cm, so that result alone does not demonstrate an improvement.
NEXT EXPERIMENT / PROPOSED
Source: my Day 1 & Day 2 experiment logs. The videos above are project demonstrations; no before/after pairing is asserted.
03HARDWARE / CONTROL
Unitree GO-8010-6 · Joint actuator control
A hands-on actuator platform for exploring communication, motor feedback, and joint-level motion control.
Established RS485 communication, implemented position, velocity, and torque command interfaces in C++, and read actuator feedback. Project work also includes single-joint dynamics modeling and LQR experiments.
04PERCEPTION / PHYSIOLOGICAL SENSING
rPPG · Real-time physiological sensing
A research-to-demo implementation of camera-based physiological signal extraction, linking video acquisition and facial region detection with real-time processing.
Implemented video acquisition, facial region-of-interest detection, and physiological signal extraction as part of rPPG algorithm integration. The recording demonstrates the interface and pipeline; quantitative validation is not included here.
CONTINUING THE EXPLORATION
05 / KINEMATICS
Robot kinematics simulations and visual studies of how joint motion relates to end-effector movement.
Visual documentation forthcoming06 / MECHANICAL DESIGN
My robotic arm assembly in SolidWorks, showing the articulated structure, joint connections, and base assembly.
ABOUT / RESEARCH DIRECTION
I am an undergraduate in Information Management and Information Systems at Beijing Jiaotong University, also pursuing a BS in Management Information Systems at Rochester Institute of Technology, with graduation expected in 2027.
My interests span reinforcement learning, imitation learning, vision-language-action models, and the hardware that makes learned behavior possible. I am preparing for graduate study in embodied intelligence.
