Model Adaptation for Time Constrained Embodied Control

TL;DR

MoDeC introduces a constraint-aware modular framework for time-constrained embodied control, framing multi-task RL with latency bounds and deploying a three-component architecture: a modular base network, a module-selection network, and a device adapter. The approach leverages soft modularization, iterative module selection, knowledge distillation, and few-shot adaptation to dynamically allocate computation and meet device-specific latency constraints. Empirical results across Meta-world, CARLA, and AI2THOR on Nvidia Jetson platforms show MoDeC outperforms dynamic baselines in both success rate and constraint adherence, including improvements of up to 14.4% over DS-Net in CARLA and robust unseen-task performance in AI2THOR. These findings demonstrate that constraint-aware modular routing can enable reliable, real-time multi-task control in diverse embodied settings with heterogeneous hardware.

Abstract

When adopting a deep learning model for embodied agents, it is required that the model structure be optimized for specific tasks and operational conditions. Such optimization can be static such as model compression or dynamic such as adaptive inference. Yet, these techniques have not been fully investigated for embodied control systems subject to time constraints, which necessitate sequential decision-making for multiple tasks, each with distinct inference latency limitations. In this paper, we present MoDeC, a time constraint-aware embodied control framework using the modular model adaptation. We formulate model adaptation to varying operational conditions on resource and time restrictions as dynamic routing on a modular network, incorporating these conditions as part of multi-task objectives. Our evaluation across several vision-based embodied environments demonstrates the robustness of MoDeC, showing that it outperforms other model adaptation methods in both performance and adherence to time constraints in robotic manipulation and autonomous driving applications

Figures (5)

• Figure 1: Overall Architecture
• Figure 2: Learning Procedure of $\textsf{MoDeC}$. On the left side of the figure, the base network and the iterative module selection network are jointly leaned through a reward function $R_\text{ims}$. The iterative module selection network then distills into a single-step decision module selection network, as shown in the middle side. Finally, as depicted on the right side, the device adapter utilizes few-shot samples to associate the inference time with the number of modules (module utilization), effectively transforming the constraint representation into a specific number of modules to use for different devices.
• Figure 3: Environments
• Figure 4: Effect of Distillation
• Figure 5: Effect of Base Network Architecture