Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bridging the Resource Gap: Deploying Advanced Imitation Learning Models onto Affordable Embedded Platforms

Haizhou Ge, Ruixiang Wang, Zhu-ang Xu, Hongrui Zhu, Ruichen Deng, Yuhang Dong, Zeyu Pang, Guyue Zhou, Junyu Zhang, Lu Shi

TL;DR

A pipeline that facilitates the migration of advanced imitation learning algorithms to edge devices is proposed via an efficient model compression method and a practical asynchronous parallel method Temporal Ensemble with Dropped Actions (TEDA) that enhances the smoothness of operations.

Abstract

Advanced imitation learning with structures like the transformer is increasingly demonstrating its advantages in robotics. However, deploying these large-scale models on embedded platforms remains a major challenge. In this paper, we propose a pipeline that facilitates the migration of advanced imitation learning algorithms to edge devices. The process is achieved via an efficient model compression method and a practical asynchronous parallel method Temporal Ensemble with Dropped Actions (TEDA) that enhances the smoothness of operations. To show the efficiency of the proposed pipeline, large-scale imitation learning models are trained on a server and deployed on an edge device to complete various manipulation tasks.

Bridging the Resource Gap: Deploying Advanced Imitation Learning Models onto Affordable Embedded Platforms

TL;DR

A pipeline that facilitates the migration of advanced imitation learning algorithms to edge devices is proposed via an efficient model compression method and a practical asynchronous parallel method Temporal Ensemble with Dropped Actions (TEDA) that enhances the smoothness of operations.

Abstract

Advanced imitation learning with structures like the transformer is increasingly demonstrating its advantages in robotics. However, deploying these large-scale models on embedded platforms remains a major challenge. In this paper, we propose a pipeline that facilitates the migration of advanced imitation learning algorithms to edge devices. The process is achieved via an efficient model compression method and a practical asynchronous parallel method Temporal Ensemble with Dropped Actions (TEDA) that enhances the smoothness of operations. To show the efficiency of the proposed pipeline, large-scale imitation learning models are trained on a server and deployed on an edge device to complete various manipulation tasks.

Paper Structure

This paper contains 14 sections, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Embodied Intelligence System
  4. Edge AI
  5. Method
  6. Model Compression
  7. Temporal Ensemble with Dropped Actions (TEDA)
  8. Experiment
  9. Manipulation Platforms and Tasks
  10. Data Collection and Policy Training
  11. Experiment Results
  12. With/Without ISU
  13. Overall Task Results
  14. Limitations and Discussions

Figures (3)

  • Figure 1: Pipeline for deploying Imitation Learning algorithms on embedded devices. The policy trained with collected data is deployed into the embedded device through two key components: model compression and Temporal Ensemble with Dropped Actions (TEDA) with a trunking size of 5 for policy inference ($k=5$). In the TEDA section, the blue block represents the normal action, while the black block represents the dropped action and the red box represents Temporal Ensemble when applying actions.
  • Figure 2: Overview of teleoperation systems for data collection. Left: Single-arm with a two-finger gripper. Middle: Single-arm with a three-finger gripper. Right: Dual-arm with grippers.
  • Figure 3: Task 1: Single-Arm Cup Stacking:This task involves using a single robotic arm with a two-finger gripper to place a blue cup into a pink cup. First, the robotic arm approaches the blue cup and picks it up (Subtask 1: Grasp). Then, the arm adjusts the position of the blue cup to place it into the pink cup. Finally, the arm opens the gripper and returns to the initial position (Subtask 2: Place).