Project Database

This project has been taken

Machine learning has shown great potential for enabling robots to acquire robust and adaptive locomotion skills. For hyper-redundant robots, such as salamander-inspired robots, this remains challenging because of the high-dimensional body dynamics, the diversity of possible behaviors, and the need to integrate multimodal sensory feedback from both terrestrial and aquatic environments.

This project is an extension of a previous project, in which we will keep exploring learning-based methods for improving salamander robot locomotion in complex amphibious environments. Possible directions include sensor fusion, maneuvering, transition, sim-real transfer, etc. A key goal will be to evaluate whether biological/physical priors can improve learning efficiency, robustness, and smooth transitions between behaviors.

The project is suitable for a highly self-motivated student with complete project experience in machine learning. Familiarity with MuJoCo MJX or other physics simulators is expected. Experience with CPGs, signal processing, reinforcement learning, or sensor-based control will be helpful.

Students who are interested in this project shall send the following materials to the assistant: (1) resume, (2) transcript showing relevant courses and grades, and (3) other materials that can demonstrate their skills and project experience (such as videos, slides, Git repositories, etc.).

This project is part of an ongoing collaboration with Prof. Eva Naumann at Duke University (USA) to study and model zebrafish sensorimotor integration using simulation and robots. It follows two collaborative studies that were published in Science robotics, see Artificial embodied circuits uncover neural architectures of vertebrate visuomotor behaviors, Energy efficiency and neural control of continuous versus intermittent swimming in a fishlike robot and FARMS: Framework for Animal and Robot Modeling and Simulation.

Depending on the interest of the student, there are several new avenues to be explored. These include: (i) improving the neuromechanical simulation and integrating it into a new simulation framework (FARMS, link), (ii) applying deep reinforcement learning for designing biologically constrained sensory-motor circuits, (iii) investigating new visually-guided behaviors, and/or (iv) programming and improving ZBot for neuroscience related studies.

The project will normally take place at EPFL, but could also happen at Duke University if there is interest.

In nature, animals have many compliant structures that benefit their locomotion. For example, compliant foot/leg structures help adapt to uneven terrain or negotiate obstacles, flexible tails allow efficient undulatory swimming, and muscle-tendon structures help absorb shock and reduce energy loss. Similar compliant structures may benefit salamander-inspired robots as well.

In this study, the student will try simulating compliant structures (the feet of the robot) in Mujoco and optimizing the design. To bridge the sim-to-real gap, the student will first work with other lab members to perform experiments to measure the mechanical properties of a few simple compliant structures. Then, the student needs to simulate these experiments using the flexcomp plugin of Mujoco or theoretical solid mechanics models, and tune the simulation models to match the dynamical response in simulation with the experiments. Afterward, the student needs to optimize the design parameters of the compliant structures in simulation to improve the locomotion performance of the robot while maintaining a small sim-to-real gap. Finally, prototypes of the optimal design will be tested on the physical robot to verify the results.

The student is thus required to be familiar with Python programming, physics engines (preferably Mujoco), and optimization/learning algorithms. The student should also have basic mechanical design abilities to design mechanical structures and perform experiments. Students who have taken the Computational Motor Control course or have experience with data-driven design and solid mechanics would also be preferred.

The student who is interested in this project shall send the following materials to the assistants: (1) resume, (2) transcript showing relevant courses and grades, and (3) other materials that can demonstrate your skills and project experience (such as videos, slides, Git repositories, etc.).

Animals display a rich diversity of behaviors in natural environments, but only a small set of them have been well studied and simplified into template gaits. Recent advances in imitation learning have enabled quadruped and humanoid robots to reproduce complex animal and human motions, but these approaches have been less explored for amphibious robots that must coordinate body, limbs, and environmental interactions across water and land.

In this project, we will work on the recordings of real salamanders in a vivarium with various terrains. We will extract multi-terrain locomotion behaviors from the collected motion data. Then we will investigate how diverse salamander behaviors can be transferred to simplified salamander models or salamander-inspired robots. Note that because we rely on DeepLabCut to track the animal kinematics from videos, the student may spend half a semester on labeling video data , unless a better tracking method is found or developed.

The student is expected to be familiar with Python programming, signal processing, kinematics, and imitation learning. Experience with dimensionality reduction, MuJoCo simulation, and CPGs is highly recommended.

Students who are interested in this project shall send the following materials to the assistants: (1) resume, (2) transcript showing relevant courses and grades, and (3) other materials that can demonstrate their skills and project experience (such as videos, slides, Git repositories, etc.).

This project has been taken for 2026 Fall semester.

INTRODUCTION

Generalist Vision-Language-Action (VLA) policies can now perform a wide range of robot manipulation skills, but evaluating and improving them remains slow and expensive: rigorous evaluation requires hundreds of real-world rollouts, and systematic improvement demands additional corrective data with expert labels. Generative world models offer a scalable alternative by letting policies roll out inside imagination space, and recent work has shown that controllable, multi-view, action-conditioned world models can both rank policy performance without real-world execution and boost success rates by synthesising successful trajectories for supervised fine-tuning. In parallel, goal-conditioned formulations specify tasks through a target visual or language goal rather than a scalar reward, enabling reward-free online improvement through hindsight relabelling.

OBJECTIVES

This project will tightly couple these three directions on our existing manipulation platforms at EPFL. Concretely, we will:

1. Integrate a controllable multi-view world model with our ROS2-based robot platforms (ViperX 300S and WidowX-250 arms, optionally a humanoid such as Reaman), including time-synchronisation with RGB-D cameras and robot state.
2. Adapt the world model into a goal-conditioned predictor that, given a current observation and a visual or language goal, imagines temporally consistent multi-view futures.
3. Implement an iterative co-improvement loop in which a small batch of real-world policy rollouts is used to ground the world model on contact-rich, deformable-object tasks, and the grounded model in turn generates large-scale synthetic data.
4. Replace the explicit success-classifier reward model with hindsight goal relabelling and LoRA-based fine-tuning of a π0.5-class VLA policy, enabling reward-free autonomous improvement.
5. Perform systematic evaluation, ablation, and documentation to deliver a reusable goal-conditioned RL pipeline for future projects.

IMPORTANCE

We have well-documented tutorials on using the robots, teleoperation interfaces for data collection, the HPC cluster, and a complete pipeline for training robot policies. Open-source codebases for π0.5 (openpi), Ctrl-World, Act2Goal, and video diffusion will be integrated into this ecosystem, so the student can focus on the research questions rather than low-level setup.

We already have a strong base and results from an ongoing project: world model and RL for robot manipulation. Compared to standard language-conditioned VLAs trained only on demonstrations, this paradigm gives us:

• Imagination-based policy evaluation that closely tracks real-world performance rankings, with no physical rollouts.
• Targeted policy improvement on unseen objects, novel spatial layouts, and out-of-distribution instructions via synthetic successful trajectories.
• Physically grounded predictions of contact-rich and deformable-object interactions once the world model is fine-tuned on online rollout data.
• Reward-free online adaptation through hindsight relabelling.

WHAT WE HAVE

1. Ready-and-easy-to-use robot platforms: ViperX 300S and WidowX-250 arms configured with 4 RealSense D405 cameras, various grippers, and a mobile robot platform, fully compatible with the LeRobot V3.0 data format.
2. Pretrained models: access to π0.5 / π0-FAST policies, the Ctrl-World world-model checkpoint, and Stable Video Diffusion backbones.
3. Computing resources: two desktop PCs with NVIDIA GPUs 5090 and 4090, plus HPC cluster access for world-model fine-tuning.

CANDIDATES

Interested students can apply by emailing sichao.liu@epfl.ch or lixuan.tang@epfl.ch. Please attach your transcript and a short description of your past/current experience on related topics such as robotics, computer vision, reinforcement learning, generative models, and VLA.

The position is open until we have final candidates. Otherwise, the position will be closed.

RECOMMENDED READING

1. “Ctrl-World: A Controllable Generative World Model for Robot Manipulation.” ICLR 2026. arXiv:2510.10125.
2. “VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model.” arXiv preprint arXiv:2602.12063, 2026.
3. “Act2Goal: From World Model To General Goal-conditioned Policy.” arXiv preprint arXiv:2512.23541, 2025.
4. “LeRobot: An Open-Source Library for End-to-End Robot Learning.” arXiv preprint arXiv:2602.22818, 2026.
5. “π0.5: A Vision-Language-Action Model with Open-World Generalization.” arXiv preprint arXiv:2504.16054, 2025.
6. “Evaluating Robot Policies in a World Model.” arXiv preprint arXiv:2506.00613, 2025.

This project has been taken for 2026 Fall semester.

Background

To improve the indoor autonomy of people with mobility impairments, including wheelchair users, everyday furniture can be augmented with mobility and manipulation capabilities. Such robotic furniture could autonomously reorganize room layouts, clear pathways, transport objects, and grasp and deliver everyday items to users.

This master’s thesis aims to integrate mobile robotic furniture, an origami-inspired robotic arm, markerless perception, and language-based interfaces into a complete assistive mobile manipulation system.

Current Progress

1. Mobile robotic furniture: We have developed a group of mobile robotic furniture units that can autonomously navigate and reconfigure indoor environments.
   References: Publication 1, Publication 2


2. Multi-robot dense coordination: Our Velocity Potential Field Modulation (VPFM) algorithm guides multiple robotic furniture units toward their target poses while avoiding collisions and deadlocks.
   Reference: VPFM publication


3. Markerless tracking and body-language interaction: A multi-view RGB-D system tracks users and furniture units in real time. It provides visual feedback for multi-robot coordination and enables users to control robotic furniture through simple hand motions. The single-view version has been published, while the multi-view version is currently under review.
   References: Single-view system, AI for Good Global Summit 2026 demonstration


4. LLM/VLM supported language interfaces: A natural-language interface supported by a large language model and a vision-language model allows the user to request tasks such as “Bring me a cup of water.” The system can interpret the request, localize the desired object, and identify the furniture on which it is placed. The related paper is under preparation, and a demonstration can be presented during the interview.


5. Origami-inspired robotic arm: The project uses an origami-inspired arm developed by the EPFL Reconfigurable Robotics Lab. The arm can remain folded when not in use and extend for grasping tasks, making it suitable for integration with robotic furniture.
   Reference: Origami-inspired third arm

Thesis Objectives

1. Mechanical and electronic integration: Integrate the origami-inspired arm with a mobile robotic furniture platform. An initial configuration is available, but the student will be encouraged to improve the design to increase the workspace, reachability, and stability of the mobile manipulator.


2. Whole-body control: Develop a controller that coordinates the mobile base and robotic arm. The controller will use feedback from the markerless tracking system and the VLM-based object-grounding module to generate commands for positioning, navigation, grasping, and collision avoidance.


3. End-to-end demonstration: Implement the following scenario:
   • A wheelchair user says, “Bring me a cup of water.”
   • The system interprets the request and localizes the target cup on a static piece of furniture.
   • The system may ask the user for confirmation when the request is ambiguous.
   • A mobile table equipped with the origami arm navigates to the object and positions itself for manipulation.
   • The arm grasps the cup, and the mobile table brings it back to the user.

The final demonstration will integrate speech interaction, visual grounding, markerless tracking, mobile navigation, whole-body control, and robotic grasping.

Desired Candidate Profile

• Experience with ROS 2, C++, and Python;
• Experience with mechanical design or rapid prototyping;
• Interest in mobile manipulation, assistive robotics, robot perception, or human–robot interaction;
• Experience developing real-world robotic applications is an advantage.

This master thesis will be co-supervised by BioRob and RRL

• Contact in BioRob: Lixuan Tang (lixuan.tang@epfl.ch)
• Contact in RRL: Alexander Schuessler (alexander.schuessler@epfl.ch)

This project has been taken for 2026 fall semester

Background

To improve the indoor autonomy of people with mobility impairments, including wheelchair users, everyday furniture can be equipped with mobility and manipulation capabilities. Such robotic furniture could autonomously reorganize room layouts, clear pathways, transport objects, and grasp and deliver everyday items to users.

Coordinating multiple robotic furniture units in confined indoor environments is challenging because their physical shapes and dimensions cannot be neglected. Unlike point robots, furniture units must account for their full geometry when navigating through narrow spaces and interacting with one another.

Current Progress

Our previously developed algorithm, Velocity Potential Field Modulation (VPFM), has demonstrated strong performance in the dense coordination of polytopic robot swarms in 2D environments. Given the current pose and velocity of each robot, together with its target pose, VPFM generates collision-aware motion commands that guide all robots toward their assigned targets while reducing the risk of collisions and deadlocks.

More information is available in our IEEE Robotics and Automation Letters publication and the corresponding open-source repository:

Paper: IEEE RA-L publication

Project Objectives

1. Adapting VPFM to an indoor maze map with static obstacles.
   An ordinary home environment consists of multiple rooms separated by walls and doors, as well as static furniture pieces. To ensure that each mobile robotic furniture unit can arrive at any reachable position in the home environment, combining VPFM with a global path planner can help the swarm of mobile robotic furniture to overcome bottleneck regions and resolve the deadlock.


2. Conduct simulation-based validation.
   Implement and evaluate the proposed methods in simulation using representative dense indoor scenarios. The evaluation should investigate collision avoidance, deadlock resolution, convergence to target poses, computational efficiency, and the benefits of combining a global path planner with the local VPFM coordinator.

This project has been taken for 2026 fall semester

Background

To improve the indoor autonomy of people with mobility impairments, including wheelchair users, everyday furniture can be equipped with mobility and manipulation capabilities. Such robotic furniture could autonomously reorganize room layouts, clear pathways, transport objects, and grasp and deliver everyday items to users.

Coordinating multiple robotic furniture units in confined indoor environments is challenging because their physical shapes and dimensions cannot be neglected. Unlike point robots, furniture units must account for their full geometry when navigating through narrow spaces and interacting with one another.

Current Progress

Our previously developed algorithm, Velocity Potential Field Modulation (VPFM), has demonstrated strong performance in the dense coordination of polytopic robot swarms in 2D environments. Given the current pose and velocity of each robot, together with its target pose, VPFM generates collision-aware motion commands that guide all robots toward their assigned targets while reducing the risk of collisions and deadlocks.

More information is available in our IEEE Robotics and Automation Letters publication and the corresponding open-source repository:

Paper: IEEE RA-L publication

Project Objectives

1. Extend VPFM to height-aware, multi-layer coordination.
   Develop a representation and coordination framework that accounts for the three-dimensional geometry of robotic furniture while maintaining predominantly planar motion. This extension should allow compatible parts of different furniture units to overlap vertically. For example, a low table may move underneath a higher table, and the seat of a chair may pass below a tabletop, while the chair back must still be treated as a potential collision.


2. Conduct simulation-based validation.
   Implement and evaluate the proposed methods in simulation using representative dense indoor scenarios. The evaluation should investigate collision avoidance, deadlock resolution, convergence to target poses, computational efficiency, and the benefits of height-aware coordination compared with the original two-dimensional formulation.