Scholarlink

<p>Table tennis with collaborative robots has been a challenge in robotics for decades, due to its unique challenges, primarily requiring precise real-time ball trajectory predictions to enable responsive, accurate gameplay. Traditional physical models, while widely studied, struggle with factors like spin, limiting accuracy. With advancements in computational power, data-driven approaches as sequence-to-sequence (seq2seq) models as trajectory autoencoders (TAE) offer potential for improved long-term prediction. However, they are not fully optimized for time-series prediction. More advanced seq2seq models with recurrent layers are better suited and improve prediction accuracy, but remain underexplored for trajectory prediction. Additionally, recurrent layers integrated with TAE, as gated-recurrent unit TAE (GRU-TAE), have not been applied for real-world trajectories. This study introduces a novel GRU-TAE model designed for high-accuracy, low-latency trajectory predictions in table tennis. Our approach was evaluated on both real and simulated data, demonstrating that TAE-based architectures outperform traditional recurrent models (e.g., LSTM, GRU, RNN) by 44% in long-term prediction accuracy, achieving an average computation time of 8.2 ms. Additionally, GRU-TAE improves accuracy by 14% over traditional TAE models and by 41% compared to baseline model-based methods in predicting real ball trajectories. These results are important because they enhance robotic arm performance in returning balls to human players, allowing earlier positioning for ball return, and reducing late position adjustments. This work sets the foundation for integrating GRU-TAE with advanced model-based approaches, aiming toward real-world deployment in collaborative table tennis robots.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-time trajectory prediction of a ping-pong ball using a GRU-TAE

  • 期刊论文

  • 发表日期：

    2025年1月17日

查看全文
  • Baptiste Toussaint,
  • Maxime Raison

摘要

Table tennis with collaborative robots has been a challenge in robotics for decades, due to its unique challenges, primarily requiring precise real-time ball trajectory predictions to enable responsive, accurate gameplay. Traditional physical models, while widely studied, struggle with factors like spin, limiting accuracy. With advancements in computational power, data-driven approaches as sequence-to-sequence (seq2seq) models as trajectory autoencoders (TAE) offer potential for improved long-term prediction. However, they are not fully optimized for time-series prediction. More advanced seq2seq models with recurrent layers are better suited and improve prediction accuracy, but remain underexplored for trajectory prediction. Additionally, recurrent layers integrated with TAE, as gated-recurrent unit TAE (GRU-TAE), have not been applied for real-world trajectories. This study introduces a novel GRU-TAE model designed for high-accuracy, low-latency trajectory predictions in table tennis. Our approach was evaluated on both real and simulated data, demonstrating that TAE-based architectures outperform traditional recurrent models (e.g., LSTM, GRU, RNN) by 44% in long-term prediction accuracy, achieving an average computation time of 8.2 ms. Additionally, GRU-TAE improves accuracy by 14% over traditional TAE models and by 41% compared to baseline model-based methods in predicting real ball trajectories. These results are important because they enhance robotic arm performance in returning balls to human players, allowing earlier positioning for ball return, and reducing late position adjustments. This work sets the foundation for integrating GRU-TAE with advanced model-based approaches, aiming toward real-world deployment in collaborative table tennis robots.