PhD in Physical AI for Adaptive Robot Manipulation

Aarhus University (AU); Aarhus; On-site; DKK 400,000 - 450,000; Full time

Job summary

Aarhus University invites applications for a fully funded three-year PhD position in Physical AI for Adaptive Robot Manipulation. The role is based in the Adaptive & Agentic AI (A3) Lab at the ECE department, incl. collaboration with SDU. The project STORM aims to give robotic foundation models a sense of touch.

The candidate will build the ML core, study contact representations, and validate methods on dexterous manipulation with safety constraints.

Qualifications

• Master's degree in a quantitative field with start date before or at the begin date of the PhD.
• Strong background in ML, deep learning, CV or robotics with motivation in Physical AI.
• Experience with Python/PyTorch and modern deep architectures.
• Hands-on robotics experience or willingness to work with hardware.

Responsibilities

• Develop the machine-learning core of STORM, including tactile sensing approaches.
• Investigate contact representations and short-horizon predictive models.
• Explore parameter-efficient adaptation of foundation policies while ensuring safety.
• Evaluate methods on public benchmarks and real robotic setups.

Skills

Python

PyTorch

Transformers

Machine Learning

Robotics

Vision-Language Models

English communication

Education

Master's degree in CS/ML/Robotics

Tools

ROS 2

Job description

Title

PhD in Physical AI for Adaptive Robot Manipulation

Research area and project description

Applications are invited for a fully funded three-year PhD position at the Department of Electrical and Computer Engineering, Aarhus University. The position is part of STORM (Soft Tactile Memory for Safe Physical-AI Correction of Robotic Foundation Models), a new project funded by the Independent Research Fund Denmark (DFF). The candidate will join the Adaptive & Agentic AI (A3) Lab, headed by Dr. Behzad Bozorgtabar (main supervisor), and will work closely with Professor Ahmad Rafsanjani and the SDU Soft Robotics group, who are developing the soft tactile skin.

Research vision

How can a robot foundation model use physical feedback to correct an action during execution? Models such as vision-language-action (VLA) and world-action models let robots follow natural-language instructions, but they act on what they see rather than on what they feel. When a grasp begins to slip or a fragile object deforms, the camera view often looks fine until the damage is done. STORM gives robotic foundation models a sense of touch, so that a robot can notice when a manipulation is going wrong and correct itself safely while it acts.

PhD project

The candidate will develop the machine-learning core of STORM. The work spans three connected scientific questions, leaving room for the candidate to shape the specific algorithmic directions: (1) How can contact representations be learned from advanced tactile interfaces, including a soft-fluidic tactile skin developed at SDU? (2) How can predictive models forecast how contact states should evolve over short horizons under a policy's intended actions? (3) How can the mismatch between expected and felt contact drive parameter-efficient adaptation of pre-trained foundation policies while guaranteeing safety? Methods will be evaluated on public benchmark and real-world robotic setup, focusing on dexterous manipulation of fragile and unseen objects as well as safe human-robot interaction.

Environment

The project sits at the intersection of machine learning, Physical AI, robot foundation models (e.g., VLA and World-Action models), world models, tactile sensing, and adaptation algorithms—areas in which the A3 Lab publishes regularly at top-tier venues such as NeurIPS, ICLR, CVPR, and CoRL. The PhD includes a possible planned research stay at KTH Royal Institute of Technology in Stockholm, and close collaboration with the SDU team.

Qualifications and specific competences

• Education: A master’s degree (120 ECTS or equivalent) in Computer Science, Machine Learning, Artificial Intelligence, Robotics, Electrical/Computer Engineering, or a related quantitative field, completed or nearing completion before the start date.
• Technical Skills: Advanced proficiency in Python and PyTorch, with hands‑on experience training, fine‑tuning, or adapting modern deep architecture such as Transformers, state‑space/sequence models, and diffusion or flow‑matching models.
• Core Knowledge: A strong foundation in machine learning, deep learning, or computer vision, combined with a strong motivation to work on Physical AI, multimodal tactile sensing, world models, and online adaptation, etc.
• Relevant Research Experience:
  • Robot & Policy Learning: Imitation learning, behavior cloning, offline/model‑based RL, or policy learning for manipulation (e.g., diffusion policies, Vision‑Language‑Action models, or World‑Action models), as well as familiarity with manipulation benchmarks (e.g., LIBERO, CALVIN, or ManiSkill).
  • World Models & Sequential Representation Learning: Self‑supervised learning on temporal or multimodal sensor streams, action‑conditioned video/dynamics prediction, latent world models, or uncertainty estimation.
  • Model Adaptation & Post‑Training: Test‑time adaptation (TTA), continual/online learning, parameter‑efficient fine‑tuning (PEFT/LoRA), or safe/constrained policy updates.
• Hands‑on Robotics: Prior experience with physical robot platforms, grippers, tactile/force sensors, and ROS 2 is an advantage, but not strictly required for candidates with an exceptional machine learning background. However, a strong willingness to work with physical hardware and run real‑robot experiments is required.
• Track Record: Evidence of strong research potential—such as a strong master’s thesis, open‑source research code, or publications/submissions at leading ML, vision, or robotics venues (e.g., NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, CoRL, RSS, ICRA)—is a strong advantage.
• Attributes: A mindset for scientific rigor, reproducibility, and open‑source contribution, the ability to collaborate across algorithmic AI and physical hardware teams, and excellent written and spoken English.

Application Requirements

• Statement of Interest (1 page): Detailing your background in machine learning, computer vision or robotics, any work experience relevant to the position, and your motivation for joining the A3 Lab and the project.
• Curriculum Vitae: Including a publication list (if applicable) and a technical project page, with links to code repositories or videos of experiments where available.
• Academic Records: Transcripts and diplomas (Bachelor’s and Master’s).
• Project Description Requirement: The candidate is required to use the project description from the PhD announcement.

Application Deadline

04 January 2027 at 23:59 CET

Preferred starting date

01 March 2027

Place of employment and place of work

The place of employment is Aarhus University, and the place of work is Adaptive & Agentic AI (A3) Lab, Department of Electrical and Computer Engineering (ECE), Faculty of Technical.

Contacts

• Behzad Bozorgtabar, e‑mail: [e-mail] (main supervisor)
• [e-mail]

Formalities and salary range

Salary and terms as agreed between the Danish Ministry of Taxation and the Confederation of Professional Unions. You will be employed as an academic staff member at basic salary grades 4–8. As a PhD fellow you will also receive a monthly non‑pensionable allowance.

Equal opportunities statement

Aarhus University’s ambition is to be an attractive and inspiring workplace for all and to foster a culture in which each individual has opportunities to thrive, achieve and develop. We view equality and diversity as assets, and we welcome all applicants.