Mr. Clean
A robotics system combining collision-free navigation, reinforcement-learning object pickup, and multi-camera manipulation trajectories.
NAVIGATION / REINFORCEMENT LEARNING / MANIPULATION
Three ways to control the robot
Click the floor to add boxesRobotics stack
Simulated lidar, wheel encoders and IMU feed SLAM Toolbox in ROS 2. Obstacles are inflated by the robot’s footprint, A* finds a collision-free route, and ACT predicts the arm’s reach, grasp and lift as 32-step action chunks.
Robotics stack · 0 steps travelled
Drop boxes on the floor, even while the robot is driving, and it plans a new route from where it stands. The routes are computed in your browser for illustration; the real controllers run in MuJoCo, and the fly-brain trail is a stand-in for its reactive behaviour.
- Collision-free navigation routes
- 9/9
- Object pickup success
- 95%
About this project
Built an A* planner with obstacle inflation, spline trajectories, LiDAR and odometry, and closed-loop control. Achieved nine out of nine collision-free navigation routes with zero falls and docking error below 10 cm and 5°.
Trained a reinforcement-learning policy in MuJoCo for household objects, achieving 95% pickup success.
Implemented an Action Chunking Transformer that fuses three camera streams with robot state through ResNet-18, a CVAE encoder, and a transformer to generate 32-step manipulation trajectories.
How it works
- PERCEPTIONLiDAR, odometry + three cameras
- NAVIGATIONA*, spline paths + docking
- MANIPULATIONRL pickup + 32-step ACT
Built with
- Python
- PyTorch
- MuJoCo
- ROS 2
- OpenCV
- Stable-Baselines3
- Docker