Scholarlink

This study proposes a curriculum learning framework to evolve voxel-based soft robot morphologies and optimize their control to solve multiple tasks. Task mismatch as well as the catastrophic forgetting during control optimization present a significant challenge to realizing versatile robots across tasks. This work proposes a curriculum learning approach with sequential task learning and periodic task rotation to optimize the control of a given morphology for multiple tasks. Three EvoGym tasks viz. Walker-v0, BridgeWalker-v0 and DownStepper-v0 have been chosen for the MorphoEvolution experiments. A two as well as three EvoGym task combination scenarios have been used to evaluate the effectiveness of the proposed framework. The evolved morphologies demonstrated appropriate control optimized for the chosen scenario. The curriculum learning framework has been shown to demonstrate the potential for multi-task morphological evolution.

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

From Specialists to Generalist Soft Robots: Evolving a Single Morphology and Controller Across Tasks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • V. Sabarish,
  • C. Shunmuga Velayutham

摘要

This study proposes a curriculum learning framework to evolve voxel-based soft robot morphologies and optimize their control to solve multiple tasks. Task mismatch as well as the catastrophic forgetting during control optimization present a significant challenge to realizing versatile robots across tasks. This work proposes a curriculum learning approach with sequential task learning and periodic task rotation to optimize the control of a given morphology for multiple tasks. Three EvoGym tasks viz. Walker-v0, BridgeWalker-v0 and DownStepper-v0 have been chosen for the MorphoEvolution experiments. A two as well as three EvoGym task combination scenarios have been used to evaluate the effectiveness of the proposed framework. The evolved morphologies demonstrated appropriate control optimized for the chosen scenario. The curriculum learning framework has been shown to demonstrate the potential for multi-task morphological evolution.