Robust Motion Mapping Between Human and Humanoids Using CycleAutoencoder

Teleoperation needs accurate and robust motion mapping between human and humanoid motion to generate intuitive robot control with human-like motion. Data-driven methods are often deployed as it can result in intuitive, real time motion mapping. When using these methods, the common focus is on the ac...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2021 IEEE International Conference on Robotics and Biomimetics (ROBIO) s. 93 - 98
Hlavní autoři: Stanley, Matthew, Tao, Lingfeng, Zhang, Xiaoli
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 27.12.2021
Témata:
Humanoid robots
Predictive models
Real-time systems
Robot control
Robustness
Training
Training data
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Teleoperation needs accurate and robust motion mapping between human and humanoid motion to generate intuitive robot control with human-like motion. Data-driven methods are often deployed as it can result in intuitive, real time motion mapping. When using these methods, the common focus is on the accuracy of the motion mapping model. However, effort needs to be put into making the mapping model robust in face of noisy or incomplete dataset. In other words, the model needs to learn the generalizable mapping rules, not just be accurate in predicting the training data. To create a robust and accurate model for motion mapping, we developed the novel CycleAutoencoder method. This method simultaneously trains two autoencoders using traditional losses, mixed losses, and cycle losses. These losses allow the autoencoders to reconstruct the motion mutually between humans and humanoids. This allows the method to learn the mapping with improved accuracy and robustness compared to training a traditional autoencoder. The results of human subject involved experiments demonstrated that the CycleAutoencoder method can achieve both accuracy and robustness for the mapping compared with other autoencoder-based mapping methods.
DOI:10.1109/ROBIO54168.2021.9739345