top of page

Using TensorFlow to Detect Custom Team Prop

In the world of competitive robotics, innovation and creativity are key drivers of success. The FIRST Tech Challenge (FTC) offers teams the opportunity to push the boundaries of technology through its Machine Learning toolchain (FTC-ML). Our team took on the challenge of using TensorFlow to detect a custom team prop, showcasing the power of custom machine learning models in robotics competitions.


The Challenge


Our game strategy relied heavily on detecting specific game elements, which required precise and reliable detection methods. We wanted to replace the standard white pixel detection with a custom team prop, which would not only improve detection accuracy but also gave us more points.


The Approach


We started with recording videos of our robot in action, both with and without the custom team prop. We then labeled each frame of the videos, highlighting where the team prop was present. This process, though time-consuming, was crucial for training the TensorFlow Object Detection (TFOD) model to recognize the prop accurately.


It was essential not to ignore frames without the team prop, as these "negative frames" were just as important for training the model. To expedite the labeling process, we leveraged OpenCV Object Tracking, which helped them label frames faster and more efficiently.


Once all frames were labeled, we generated a dataset and split it into training (90%) and testing (10%) sets. This ensured that the model was trained on a diverse range of data and could generalize well to new situations.


Training the Model


FTC-ML utilizes Google Tensor Processing Units (TPUs) behind the scenes for neural-net training. This powerful hardware accelerated the training process, allowing us to iterate quickly and fine-tune our model for optimal performance.


After several iterations of training and testing, our TensorFlow model became adept at detecting our custom team prop with impressive accuracy (98%). This achievement not only improved our robot's performance on the field but also demonstrated the team's dedication to pushing the boundaries.


Conclusion


Our use of TensorFlow to detect a custom team prop exemplifies the innovative spirit of the FIRST Tech Challenge. By leveraging custom machine learning models, teams can enhance their robot's capabilities and stand out in competition.


The journey from recording videos to training the TensorFlow model was challenging but rewarding. It showcased the power of machine learning in robotics and highlighted the importance of creativity and perseverance in tackling complex challenges.


 

Model performance



A low loss indicates that the model is performing well on the training data. 


It suggests that the predictions are close to the actual values, and the model is learning the underlying patterns in the data.



Regularization is a technique used to prevent overfitting, where a model performs well on the training data but fails to generalize to new, unseen data.


Low numbers indicates that the primary loss term dominates, and the model is focusing more on fitting the training data.



mAP (Mean Average Precision) provides a comprehensive assessment of the model's performance, taking into account its ability to correctly detect objects (precision) and its ability to detect most of the objects present in the dataset (recall). A higher mAP indicates a more accurate and reliable object detection model.





39 views0 comments

Recent Posts

See All

Comments


Subscribe

Subscribe to our mailing list for regular updates on news, events, insightful blogs, and free code!

Thanks for subscribing!

bottom of page