RAIN

Cloud Robotics - VLA (ongoing)

Introduction

I’m leading this new project on cloud robotics, which aims to develop both cloud-based and edge-based solutions for deploying and managing Vision-Language-Action (VLA) models. Currently, as a team of a postdoc researcher (me), several PhD students, and master students, we are working on the following tasks:

  • Designing a cloud-edge architecture / pipeline for VLA models in robotics applications.
  • Developing efficient model compression and adaptation techniques to deploy large VLA models on edge devices with limited resources.
  • Implementing real-time data processing and decision-making pipelines that leverage cloud computing power while minimizing latency.
  • Using reinforcement learning techniques to validate the safety and robustness of VLA models in dynamic environments.

We have built a prototype system that integrates cloud services with edge devices, which can run large VLA models, such as OpenVLA, on the cloud while performing inference and control on local devices. The prototype is expected to be released around early summer 2026 (March-June).

Information

This is an academic project initiated and led by me, starting from Sept. 2025.

Reference(s):

NOT YET AVAILABLE

DRL-MPC

DRL-Boosted Optimal Control for Flexible Navigation in Complex Environments

Introduction

This side project proposes an efficient approach to real-time collision-free navigation for mobile robots in complex environments (with non-convex obstacles). By integrating deep reinforcement learning with model predictive control, the aim is to achieve both collision avoidance and computational efficiency. The methodology begins with training a preliminary agent using deep reinforcement learning (DQN and DDPG), enabling it to generate actions for next time steps. Instead of executing these actions, a reference trajectory is generated based on them, which avoids obstacles present on the original reference path. Subsequently, this local trajectory is employed within an MPC trajectory-tracking framework to provide collision-free guidance for the mobile robot.

For multi-agent cases, decentralized training with decentralized execution (DTDE) is used for DRL agents, and distributed MPC is conducted for flexibility and scalability.

Information

This is a side project initiated by me and later expanded into a larger research project involving additional researchers.

Reference(s):

  1. IEEE CASE
  2. IEEE IROS

FON-AMR

Future-Oriented Navigation for Autonomous Mobile Robots

Introduction

As industrial environments become more dynamic and collaborative, Autonomous Mobile Robots (AMRs) are being deployed to work alongside humans. While hybrid human-robot settings offer improved efficiency and adaptability, they also challenge safety due to human behavior’s inherent unpredictability. This project is motivated by the goal of enabling AMRs to anticipate and react to dynamic environments by integrating learning-based prediction with optimization-based control. Specifically, it focuses on safer and more efficient future-oriented navigation, i.e., decision-making that incorporates predictive information about the near future.

Information

This is a five-year research project as the main topic of my PhD study. This project is supported by several larger research projects: ViMCoR (2019-2021), AIHURO (VINNOVA, 2023-2025), etc. The main collaborator is Volvo GTO. This project is also in collaboration with / founded by Chalmers AI Research Centre (CHAIR) and AstraZeneca.

Reference(s):

  1. PhD Thesis
  2. IEEE RA-L (arXiv)