BeginnerRMB 600–900 · 3–4 weeks
Differential-drive car + LiDAR SLAM navigation
A two-wheel differential chassis + single-line LiDAR; run through the full pipeline of mapping, localization, path planning, and obstacle avoidance.
What you learn
ROS 2 nodes and tf · LiDAR SLAM principles · Navigation stack configuration
Deliverables
- One reproducible mapping result
- An analysis of 'why it degrades in long-corridor scenarios'
Don't stop at getting the tutorial running. Change a parameter, cause one failure, and explain why — that's worth more than three more projects.
BeginnerRMB 1200–2000 · 5–6 weeks
Robotic arm visual grasping
A desktop 6-axis or 4-axis arm + RGB-D camera; close the loop of 'see - compute pose - plan - grasp'.
What you learn
Hand-eye calibration · Pose estimation · MoveIt and motion planning
Deliverables
- Grasp success-rate data (statistics over >= 50 trials)
- Failure case analysis (lighting / occlusion / pose error)
You must compute statistics; don't film just the one successful take. What interviewers most want to ask is 'why did the rest fail'.
IntermediateRMB 1500–2500 · 4–5 weeks
MCU + ROS 2 full-chain joint control
STM32 as the low-level controller (CAN or serial protocol), a ROS 2 host sending commands down, wrapped with a ros2_control hardware interface.
What you learn
Layered architecture and interface contracts · Real-time scheduling · Communication protocol design
Deliverables
- Architecture layer diagram + interface definitions
- End-to-end latency measured (GPIO timing + oscilloscope)
- Three ADRs documenting the trade-offs
This is the project software-background people should do most. The output isn't code — it's three documents proving 'you can make architecture decisions'.
IntermediateRMB 300–600 · 2 weeks
Real-time measurement deep dive
Run three comparisons on the same board: stock kernel, PREEMPT_RT, plus CPU isolation and real-time priorities; measure jitter with cyclictest.
What you learn
Real-time system principles · CPU isolation and interrupt affinity · Latency measurement methods
Deliverables
- Latency distribution comparison across the three setups
- Item-by-item analysis of jitter sources
- Measured conclusions for a 1kHz control loop
The smallest investment with the strongest persuasiveness. One jitter distribution chart beats ten pages of self-praise.
IntermediateRMB 800–1500 · 4–5 weeks
Teleoperation + motion retargeting
Use a phone IMU or camera as a simple motion-capture rig, retarget human pose to a servo gimbal or small humanoid, and achieve real-time following.
What you learn
Motion retargeting · Real-time link design · Safety fallbacks
Deliverables
- End-to-end latency data
- Retargeting mapping table
- A design that auto-returns to a safe state on over-limit or timeout
Isomorphic to the tech on fighting-robot arenas; telling this story is highly distinctive. Add an 'auto return on anomaly' layer and it becomes an engineering solution.
IntermediateRMB 2500–4000 · 8 weeks
Quadruped robot from scratch
A 12-DoF quadruped: train the gait in MuJoCo first, then transfer to the real robot.
What you learn
RL training pipeline · Domain randomization · Sim2Real debugging
Deliverables
- Simulation training curves
- Real-robot gait video
- Root-cause analysis of transfer failures
Don't chase parkour. Walking straight stably + surviving one side push is far more credible than ten falls of flashy tricks.
AdvancedRMB 2000–3000 · 6 weeks
Force-controlled grasping and impedance control
Add a force sensor to the end effector and implement constant-force pressing and compliant insertion.
What you learn
Impedance / admittance control · Force sensor calibration · Contact stability
Deliverables
- Contact force curves under different stiffnesses
- Measured stability boundary
- Data on how latency caps achievable stiffness
One of the abilities industry lacks most. Few undergrads in China can produce a measured 'latency vs achievable stiffness' curve.
AdvancedRMB 500–2000 for compute · 6–8 weeks
Embodied AI manipulation reproduction
Use LeRobot or ACT / Diffusion Policy to reproduce 'grasp a named object on voice command' on a low-cost arm.
What you learn
Imitation learning · Teleoperated data collection · Policy deployment
Deliverables
- A description of your dataset's scale and diversity
- Success-rate comparison (BC vs DP)
- Failure-mode analysis
The point isn't the model, it's the data. Whether you can explain 'why your data differs from everyone else's' decides whether this project has value.
Beginnerabout RMB 300 · 2 weeks
ESP32 wireless co-controller
The main controller keeps running control; an ESP32-S3 runs alongside: push key states to a phone via MQTT, pipe debug commands back, with OTA-capable firmware.
What you learn
Dual-core division of labor (one core for control, one for networking) · MQTT and topic design · OTA and rollback
Deliverables
- A judgment table of 'what data may go wireless and what must never'
- Measured latency distribution and packet loss of the wireless link
- A record of one recovery from a bricked firmware flash
The most valuable part of this project isn't the feature — it's being able to explain 'why control must not go wireless', using latency data you measured yourself.
Intermediateabout RMB 500 · 3 weeks
STM32 driver board
Design your own board: power supply + STM32 + one half-bridge driver + encoder interface + CAN transceiver. Once the prototypes come back, solder it, power it up, and tune a 1kHz current loop yourself.
What you learn
Schematics and the power tree · 4-layer layout and return paths · Prototyping and soldering · Staged power-up and debugging
Deliverables
- Power tree diagram + staged power-up test log
- One real defect and a V1 → V2 revision plan
- Step-response waveform of the 1kHz current loop
Always keep a V2. A board that works first try doesn't exist; explaining 'what went wrong in V1' is far more credible than 'success on the first shot'.