Ge Li - Robotics & ML

This is Ge Li (Bruce), 李戈

Ge Li (Bruce) I am a Ph.D. candidate in machine learning and robotics at the Karlsruhe Institute of Technology (KIT), Germany. My supervisors are Prof. Gerhard Neumann and Prof. Rudolf Lioutikov.

I focus on advancing the learning and representation of robotic motion policies, with a particular emphasis on their applications in imitation and reinforcement learning. I aim to improve model capacity, learning efficiency, and interpretability by leveraging state-of-the-art algorithms and tools.

In addition to my research, I serve as a teaching assistant at KIT, contributing to courses Cognitive Systems, Machine Learning, and Deep Reinforcement Learning. I also mentor several Bachelor and Master students for their research projects and theses.


Education & Experience

2020 - Present, Karlsruhe Institute of Technology, Germany, Ph.D. Candidate, Computer Science.

2015 - 2019, RWTH Aachen University, Germany, Master’s Degree, Computer-Aided Mechanical Engineering.

2011 - 2015, University of Science and Technology of China, Bachelor’s Degree, Mechanical Engineering.

2018 - 2019, Max Planck Institute for Intelligent Systems, Germany, Robot Research Intern.

2018, Alfred Kärcher SE & Co. KG, Germany, Robot Software Engineer Intern.


News

Jan. 2025, my latest work got accepted by ICLR 2025 as a Spotlight paper🔥. I will present it in Singapore 🇸🇬.

Dec. 2024, I presented my co-authored work in NeurIPS 2024 in Vancouver, Canada 🇨🇦.

Aug. 2024, one co-authored paper got accepted by CoRL 2024 🇩🇪.

May 2024, I presented my PhD work in ICLR 2024 in Vienna, Austria 🇦🇹.


Publications

TOP_ERL_ICLR25 TOP_ERL_ICLR25

TOP-ERL: Transformer-based Off-Policy Episodic Reinforcement Learning

Ge Li, Dong Tian, Hongyi Zhou, Xinkai Jiang, Rudolf Lioutikov, Gerhard Neumann, ICLR, 2025 Spotlight🔥 (top 5%).

See: Website | Code | OpenReview | arxiv

We propose a novel online off-policy RL methodology that utilizes a transformer-based critic to learn values of action sequences.



DIME DIME

DIME: Diffusion-Based Maximum Entropy Reinforcement Learning

Onur Celik, Zechu Li, Denis Blessing, Ge Li, Daniel Palenicek, Jan Peters, Georgia Chalvatzaki, Gerhard Neumann, Preprint, 2025.

See: arxiv

The SOTA diffusion-based online RL method that outperforms other cutting-edge diffusion and Gaussian policy methods.



IRIS_TT IRIS_UNITREE

IRIS: An Immersive Robot Interaction System

Xinkai Jiang, Qihao Yuan, Enes Ulas Dincer, Hongyi Zhou, Ge Li, Xueyin Li, Julius Haag, Nicolas Schreiber, Kailai Li, Gerhard Neumann, Rudolf Lioutikov, Preprint, 2025.

See: arxiv | Website

We present IRIS, an immersive Robot Interaction System leveraging Extended Reality (XR), designed for robot data collection and interaction across multiple simulators, benchmarks, and real-world scenarios



PC_RGB_ALIGN

Towards Fusing Point Cloud and Visual Representations for Imitation Learning

Atalay Donat, Xiaogang Jia, Xi Huang, Aleksandar Taranovic, Denis Blessing, Ge Li, Hongyi Zhou, Hanyi Zhang, Rudolf Lioutikov, Gerhard Neumann, Preprint, 2025.

See: arxiv

We propose a novel imitation learning method that effectively fuses point clouds and RGB images by conditioning the point-cloud encoder on global and local image tokens using adaptive layer norm conditioning, preserving both geometric and contextual information and achieving state-of-the-art performance on the RoboCasa benchmark.



VDD VDD_toy VDD_toy

Variational Distillation of Diffusion Policies into Mixture of Experts

Hongyi Zhou, Denis Blessing, Ge Li, Onur Celik, Gerhard Neumann, Rudolf Lioutikov, NeurIPS, 2024.

See: arxiv | NeurIPS | GitHub

We introduce Variational Diffusion Distillation (VDD), a novel method for distilling denoising diffusion policies into a Mixture of Experts (MoE).



Mamba_robotics Mamba_robotics Mamba_robotics Mamba_robotics

MaIL: Improving Imitation Learning with Selective State Space Models

Xiaogang Jia, Qian Wang, Atalay Donat, Bowen Xing, Ge Li, Hongyi Zhou, Onur Celik, Denis Blessing, Rudolf Lioutikov, Gerhard Neumann, CoRL, 2024.

See: arxiv | OpenReview | GitHub

We introduce Mamba Imitation Learning (MaIL), a novel imitation learning architecture that offers a computationally efficient alternative to state-of-the-art Transformer policies.



TCE TCE

Open the Black Box: Step-based Policy Updates for Temporally-Correlated Episodic Reinforcement Learning.

Ge Li, Hongyi Zhou, Dominik Roth, Serge Thilges, Fabian Otto, Rudolf Lioutikov, Gerhard Neumann, ICLR, 2024.

See: arxiv | OpenReview | GitHub

We propose a novel RL framework that integrates step-based information into the policy updates of Episodic RL, while preserving the broad exploration scope, movement correlation modeling and trajectory smoothness.



MP3, movement primitives policy

MP3: Movement Primitive-Based (Re-) Planning Policies

Hongyi Zhou, Fabian Otto, Onur Celik, Ge Li, Rudolf Lioutikov, Gerhard Neumann, in Conference on Robot Learning Workshop on Learning Effective Abstractions for Planning, Preprint 2023.

See: arxiv | Website | Poster

We enable a new Episodic RL framework that allows trajectory replanning in deep RL, which allows the agent to react with changing goal and dynamic perturbation.



ProDMP ProDMP

ProDMPs: A Unified Perspective on Dynamic and Probabilistic Movement Primitives.

Ge Li, Zeqi Jin, Michael Volpp, Fabian Otto, Rudolf Lioutikov and Gerhard Neumann, in IEEE Robotics and Automation Letters, RAL 2023.

See: Paper | Poster | GitHub | YouTube

We unified the Dynamic Movement Primitives and the Probabilistic Movement Primitives into one model, and achieved smooth trajectory generation, goal-attractor convergence, correlation analysis, non-linear conditioning, and online re-planing in one framework.



SVSL SVSL

Specializing Versatile Skill Libraries using Local Mixture of Experts

Onur Celik, Dongzhuoran Zhou, Ge Li, Philipp Becker, Gerhard Neumann, in Conference on Robot Learning, CoRL 2021.

See: Paper | OpenReview | YouTube

We developed a mixture of experts RL framework to learn versatile skills given the same task, such as forehand and backhand strikes in playing table tennis. Our method is able to assign policy experts to their corresponding context domains and automatically add or delete these experts when necessary.