Open Source Cloud Robotic Manipulation

Context

Recent breakthroughs in robotics have been fueled by advances in machine learning and the availability of large-scale datasets, mirroring the success of Large Language Models (LLMs). However, unlike LLMs that benefit from vast amounts of coherent internet-scale data, robotics research suffers from fragmented and inconsistent datasets, making systematic comparison and benchmarking challenging. To address this critical gap, researchers at KTH have developed CloudGripper (see fig. 1)—an open-source cloud robotics infrastructure that provides remote access to standardized robotic manipulation platforms, enabling researchers worldwide to test algorithms under identical conditions and contribute to a growing benchmark dataset.

Project Overview

This project aims to establish LASA’s participation in the emerging field of cloud robotics by connecting to the CloudGripper system and developing manipulation algorithms for standardized benchmark tasks. The primary objective is to compete in the ICRA 2026 Cloud Manipulation Competition, which will be held at one of the world’s premier robotics conferences. Through this project, you will implement and evaluate state-of-the-art manipulation techniques on real hardware accessed remotely, contributing both to LASA’s research portfolio and to the broader open-source robotics community.

Approach

The student will conduct a research cycle from implementation to competition participation:

• System Setup (Weeks 1-2): Set up CloudGripper API access, download and understand dataset, run baseline MuJoCo simulation
• Algorithm Implementation (Weeks 3-6): Implement behavior cloning baseline, train policy on dataset, test in simulation
• Real Robot Testing & Integration (Weeks 7-10): Deploy policy to real CloudGripper Robot (remote), collect performance metrics, compare sim vs. real performance
• Competition Participation (Week2 11-12): Optimize policy for competition tasks, submit to ICRA Cloud Manipulation track
• Documentation (Week 13): Write final report and prepare presentation Tools: Python, MuJoCo simulator, CloudGripper API, standard scientific libraries (NumPy, PyTorch optional), Docker

Tools: Python, MuJoCo simulator, CloudGripper API, standard scientific libraries (NumPy, PyTorch optional), Docker

Student gain: This project offers a unique opportunity to gain hands-on experience with cutting-edge cloud robotics infrastructure while developing practical manipulation skills. You will learn how to design, implement, and benchmark robotic control algorithms on real hardware without needing physical access to the robot. Participation in the ICRA competition will provide visibility within the international robotics community and valuable experience in presenting research at a top-tier conference setting.

Prerequisites

Required:

• Strong Python programming skills, familiarity with Git/GitHub
• Background in robotics, control theory, and machine learning (e.g., completed relevant EPFL courses)

Preferred

• Prior experience with robot simulators (MuJoCo, PyBullet, or similar)
• Basic understanding of manipulation or imitation learning, especially behavior cloning

References

[1] Zahid, M. & Pokorny, F. T. CloudGripper: An Open Source Cloud Robotics Testbed for
Robotic Manipulation Research, Benchmarking and Data Collection at Scale. in 2024 IEEE
International Conference on Robotics and Automation (ICRA) 12076–12082 (2024).