Scholarlink

In this paper, we propose a pretrained centralised model-based reinforcement learning method for collaborative grasping by a team of aerial and ground robots. Our method leverages demonstration data generated from the optimal reciprocal collision avoidance (ORCA) algorithm and PID controller to pretrain a dynamics model for the multi-agent system. This model is then utilised to optimise a centralised policy through model predictive control (MPC). We comprehensively conduct experiments under the scenarios with the load of three, four and six grasping points. The experiment results demonstrate that the proposed method outperforms model-free baselines, achieving higher success rates and faster convergence. Furthermore, the algorithm can be applied in a heterogeneous robot team for the collaborative grasping.

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

A Pretrained Centralised Model-Based Reinforcement Learning Method for Heterogeneous Collaborative Grasping

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yifan Liu,
  • Yuting Wang,
  • Kangyao Huang,
  • Qian Dong,
  • Jingyu Chen

摘要

In this paper, we propose a pretrained centralised model-based reinforcement learning method for collaborative grasping by a team of aerial and ground robots. Our method leverages demonstration data generated from the optimal reciprocal collision avoidance (ORCA) algorithm and PID controller to pretrain a dynamics model for the multi-agent system. This model is then utilised to optimise a centralised policy through model predictive control (MPC). We comprehensively conduct experiments under the scenarios with the load of three, four and six grasping points. The experiment results demonstrate that the proposed method outperforms model-free baselines, achieving higher success rates and faster convergence. Furthermore, the algorithm can be applied in a heterogeneous robot team for the collaborative grasping.