MIntNet: Rapid Motion Intention Forecasting of Coupled Human-Robot Systems With Simulation-to-Real Autoregressive Neural Networks

MIntNet: Rapid Motion Intention Forecasting of Coupled Human-Robot Systems With Simulation-to-Real Autoregressive Neural Networks
复制标题

DOI:
10.1109/lra.2023.3306646
复制
发表时间:
2023-10
影响因子:
5.2
通讯作者:
J. Atkins;HyunYong Lee
J. Atkins;HyunYong Lee
中科院分区:
计算机科学2区
文献类型:
--
作者:
J. Atkins;HyunYong Lee

文献摘要

相似文献

这封信描述了使用自回归神经网络(MIntNet)进行耦合人类机器人运动意图预测的模拟到真实训练管道的使用。仅使用关于交互任务的一般先验知识,生成大型模拟数据集,并用于训练经典自回归模型的多输出变体。该网络采用一种编码-解码的方法,在一系列时间窗口上构建耦合系统运动学的浓缩表示,并生成它们的浓缩潜在表示,以预测未来系统状态的多个序列。然后在10个真实人类受试者的交互任务数据上对该方法进行了测试,并评估了所提出网络的模拟到真实泛化性能,以及标准多层感知器、卷积和基于长短期记忆的网络的替代实现。结果表明,该网络具有较好的泛化性能,能够较好地预测低频扰动下非恒定曲率快速运动中的位置。MIntNet能够在200 $ $\text{ms}$窗口中准确预测未来的位置,误差为$3.1 \pm,在预测窗口上平均为$ $ $\text{mm}$,推理时间为$0.26 \pm 0.44$ $ $\text{ms}$。短期预测的性能更高,随着时间窗口的误差增加,$2.3 \pm 3.4$ $ $\text{mm}$在50 $ $\text{ms}$时,$2.4 \pm 4.4$ $ $\text{mm}$在100 $ $\text{ms}$时,$5.5 \pm 6.7$ $ $\text{mm}$在200 $ $\text{ms}$时。总之,这些属性允许对运动意图的敏捷预测,可用于通知辅助控制策略,以增强耦合人机系统的协作控制。
This letter describes the use of a simulation-to-real training pipeline using autoregressive neural networks (MIntNet) for coupled-human robot motion intention prediction. Using only general prior knowledge about the interaction task, a large simulation dataset was generated and used to train a multi-output variation of the classic autoregressive model. The network used an encoding-decoding method to construct condensed representations of the coupled system kinematics over a sequence of time windows and generated their condensed latent representations to predict multiple sequences of the future system states. This method was then tested on 10 real human subjects' data for the interaction task and the simulation-to-real generalization performance was evaluated for the proposed network along with alternative implementations of standard multilayered perceptron, convolutional, and long-short term memory based networks. Results show the proposed network has better generalization performance compared to the alternatives, capable of closely predicting positions during fast motion along non-constant curvatures subject to low-frequency disturbances. The MIntNet was able to accurately predict future positions in a 200 $\text{ms}$ window with errors of $3.1 \pm 4.8$ $\text{mm}$ averaged over the prediction window with inference times of $0.26 \pm 0.44$ $\text{ms}$. Performance was higher for short range predictions with errors over the time window growing as $2.3 \pm 3.4$ $\text{mm}$ at 50 $\text{ms}$, $2.4 \pm 4.4$ $\text{mm}$ at 100 $\text{ms}$, and $5.5 \pm 6.7$ $\text{mm}$ at 200 $\text{ms}$. Together these properties allow for agile predictions of motion intention that can be used to inform assistive control policies for enhanced collaborative control of coupled human-robot systems.