Smart Recycle

Smart Recycle: From Concept to Reality

The idea for Smart Recycle came from a common frustration figuring out which bin to use when disposing of waste. Misplaced items contaminate recycling, making the entire process inefficient. We asked ourselves:

🚀 What If a robotic system could automatically identify and sort waste?

With this vision in mind, we set out to build an AI-powered recycling assistant, using the LeRobot SO100M robotic arm, computer vision, and ACT neural network.

Step 1: Calibration

Before diving into waste sorting, we needed to calibrate the robotic arm and camera system for precise operation. We followed Seeed Studio’s calibration process, ensuring the camera’s field of view aligned perfectly with the robotic workspace.

For the robotic arm, we adjusted joint angles and endpoint positioning to ensure smooth and precise movements. This step was crucial, we didn't proper calibrate the arm on the first day, and wasted a lot of time.⏱️

We also calibrate a LeKiwi robot arm (Bobo!) to assist in training the data.

Step 2: Data Collection

The next challenge was waste identification. Following the Seeed Studio guide, we used a camera setup to capture thousands of images of recyclables and non-recyclables—plastic bottles, aluminum cans, paper, and organic waste. We labeled the dataset and trained with ACT.

Data collection for Bobo robot.

We've collected 50 episodes of the dataset for Bobo robot.

The checkout the visualized hugging face dataset here https://huggingface.co/spaces/lerobot/visualize_dataset?dataset=ibru%2Fbobo_the_picker&episode=36

Step 3: Training & Evaluation

The training ACT policy typically takes around 1.5 hours with H100

The first test was a mix of excitement and debugging. The AI model correctly identified most process, but the robotic arm initially struggled with precise positioning. Through iterative improvements, refining action accuracy and enhancing the visibility of bigger items, we improved efficiency and reduced errors.

Bobo Trained on ACT policy and also Pi0 policy on H100. ACT took around 1.5 hrs and Fine Tuning Pi0 policy took 9 hours on H100. Following are the Weights and Biases from wandb.

Interesting Find: Bobo Smart Recycler robot trying to learn if the object is in near distance to be picked or not.

Final Demo & Future Potential

Now, as we prepare for our final showcase, Smart Recycle is a fully functional prototype. To scale this system making recycling effortless and effective.

♻️ From a simple idea to a working prototype, Smart Recycle is a step toward cleaner and AI-powered smart cities.