Enhancing Backdrivable Parallel Robot Assistants with Safe Reinforcement Learning

Description

Motivation:
Robotics is entering a new era where physical Human–Robot Interaction plays a central role. Beyond industrial robots, there is growing demand for machines that can safely and intuitively collaborate with humans in healthcare, rehabilitation, logistics, and manufacturing. A key enabling technology is the backdrivable parallel robot assistant: a system in which the end-effector is connected to the base by several independent kinematic chains operating in parallel, driven by multiple low-friction actuators. Such robots exhibit low mechanical impedance, allowing humans to guide and move them as naturally as working alongside another person.

In Canada, these technologies are beginning to show significant promise. Backdrivable robot platforms are already used in industrial assembly line, neurorehabilitation and clinical assessment, where they support manufacturing, stroke recovery and mobility training. The increasing application offers fertile ground for deploying next generation human-friendly robot assistants, creating opportunities for students and researchers to shape how humans and machines will work together in the coming decades.

Despite these advances, backdrivable parallel robot assistants still face critical challenges:

- Changing payloads: (when grasping, carrying, or releasing objects) alter system dynamics, often requiring the human user to exert additional effort;
- Inertia compensation: when dealing with large payloads is a challenge especially in a hands-on-payload scenario;
- Maintaining backdrivability and safety: ensuring a smooth “hand-feel” even as the robot adapts to new conditions;
- Limited or delayed sensing: including non-synchronized estimation of certain signals (e.g., accelerations), which affects responsiveness and reliability.

PhD Objectives:
This 4-year PhD project aims to design and validate reinforcement learning-based controllers for backdrivable robots, with a focus on:


- Developing and extending a robot simulator for reinforcement learning training;
- Bridging learning-based control with classical model-based control for better interpretability and reliability;
- Design of adaptive load compensation controller using reinforcement learning, while preserving natural hand-feel during human interaction (e.g., low effort, minimal jitter, and stability);
- Integrating safety constraints (e.g., singularity avoidance, bounded assistive forces) through optimization-based safety filters (e.g., control barrier functions) to reinforcement learning training and execution;
- Evaluating the solution on both simulated and real backdrivable parallel robot platform, considering established industrial standards (e.g., ISO 9283, ISO 15066).


What we offer:

- An innovative applied project at the frontier of robotics, AI, and human-robot interaction;
- Access to experimental setups at the Laboratoire de robotique and collaboration with machine learning experts with the Laboratoire de vision et systèmes numériques;
- Opportunities to collaborate internationally and publish at top robotics venues and journals;
- Competitive funding and access to state-of-the-art computing and robotic equipment.

Requirements and Conditions

- Master's degree or equivalent in robotics, control, computer engineering, mechanical engineering, or applied mathematics;
- Strong background in robotics control, reinforcement learning, and optimization;
- Experience in at least two (2) of: i) model-based control and kinematics, ii) safe reinforcement learning, iii) parallel robots;
- Programming skills (Matlab, Python, C++ preferred) and familiarity with software tools (ROS, PyTorch, Matlab Simulink, etc.);
- Autonomy, creativity, and motivation to bridge theory and practice in robotics.

Please send your application files by email to hanqing.zhao@gel.ulaval.ca
with cc to clement.gosselin@gmc.ulaval.ca

Including the following files:
1. Motivation letter
2. CV
3. BSc and MSc transcripts.