COSM2IC: Optimizing Real-Time Multi-Modal Instruction Comprehension
Dulanga Weerakoon, Vigneshwaran Subbaraju, Tuan Tran, Archan Misra · IEEE Robotics and Automation Letters · 2022
Supporting real-time, on-device execution ofmulti-modal referring instruction comprehensionmodels is an important challenge to be tackled in embodied Human-Robot Interaction. However, state-of-the-art deep learning models are resource-intensive and unsuitable for real-time execution on embedded devices. While model compression can achieve a reduction in computational resources up to a certain point, further optimizations result in a severe drop in accuracy. To minimize this loss in accuracy, we propose theCOSM2ICframework, with a lightweight Task Complexity Predictor, that uses multiple sensor inputs to assess theinstructional complexityand thereby dynamically switch between a set of models of varying computational intensity such that computationally less demanding models are invoked whenever possible. To demonstrate the benefits ofCOSM2IC, we utilize a representative human-robot collaborative “table-top target acquisition” task, to curate a new multi-modal instruction dataset where a human issues instructions in a natural manner using a combination of visual, verbal, and gestural (pointing) cues. We show thatCOSM2ICachieves a 3-fold reduction in comprehension latency when compared to a baseline DNN model while suffering an accuracy loss of only$\sim$5%. When compared to state-of-the-art model compression methods,COSM2ICis able to achieve a further 30% reduction in latency and energy consumption for a comparable performance.