Robust robot planner for FOUP transport in human shared cleanrooms.

This PhD addresses this gap through a hybrid TAMP-RL approach for single-agent FOUP transport and handling, augmented by a Vision-Language Model (VLM) Reasoning Supervisor dedicated to resolving "stuck" states . The architecture will leverage existing robotic technologies developed at imec as safety proximity sensing, indoor localisation techniques and consists of:  

1. Task-and-Motion Planning (TAMP) for high-level task decomposition respecting cleanroom constraints (e.g., restricted zones, floor marking compliance, workspace sharing with humans, movable obstacles as chairs and computer stands and other robots).  

2. Reinforcement Learning (RL) for low-level reactive navigation and obstacle negotiation in dynamic spaces.  

3. VLM-Based Semantic Reasoning for stuck-state detection and recovery decision-making, explicitly answering "Can I move that object?" or "What are the humans doing?" when standard planning fails. 

When a robot becomes blocked, it does not simply replan paths; instead, it formulates targeted queries to a VLM: "Can this chair be moved safely?", "Is this equipment currently under maintenance?", or "Is the nearby operator busy or available for assistance?". The VLM acts as an expert arbiter, distinguishing between movable obstacles (chairs/trolleys) that can be cleared autonomously and sensitive infrastructure that must not be disturbed. This enables context-aware anomaly resolution, e.g., asking a human to move something if they are not busy, or moving an object if it is clearly a chair, while avoiding dangerous missteps like touching equipment in active maintenance.  

The research will be validated primarily in AR-based simulation (leveraging imec's lifelong-learning dataset for behavioral priors) and a physical mockup lab. Evaluation focuses on delivery success rates, stuck-state recovery time, replanning latency, and safety compliance. While also levering the safe navigation in human center environments of the imec prospect robot helper (ORBIT project).  
The framework bridges the gap between purely reactive navigation systems and knowledge-grounded autonomous decision-making. 

General research domain: adaptive planning and coordination, human-robot collaboration, robotic perception and autonomous systems. 

Keywords: task planning, adaptive planning, hierarchical planning, task and motion planning, human-robot interaction, dynamic environments, VLM models. 

Why this matters 

The transportation of FOUPs (Front Opening Unified Pods, standardized wafer containers) between processing stations in specialized semiconductor cleanrooms currently depends on human operators. This approach is increasingly unsustainable due to high training costs, challenging working conditions and shrinking workforces. Autonomous mobile robots offer a promising alternative, yet existing solutions frequently fail in dynamic, human-centric environments: they become blocked by unpredictable obstacles (e.g., humans or misplaced equipment) and lack the capability to recover autonomously when navigation fails. Developing this challenge, will gradually allow more complex tasks to be solved inside a cleanroom and beyond. 

What you will do  

You will investigate the underlying scientific contributions to develop and experimentally validate a robust planning architecture for autonomous FOUP transport in cleanrooms where robots must operate alongside humans, equipment and changing workflows. The research will focus on bridging the gap between long-horizon task planning and fast, safe adaptation to unexpected situations. 

You will: 

• Analyse and model FOUP transport workflows and cleanroom constraints, translating operational rules, restricted areas, floor markings, equipment states and human activities into representations that can be used by an autonomous robot planner. 

• Design a hierarchical task-and-motion planning architecture that generates transport plans at the global level while allowing the robot to adapt its execution locally when the environment or workflow changes. 

• Develop autonomous local replanning and recovery methods for situations such as blocked corridors, unexpected human activity, temporary obstacles and changes in the availability of transport routes or equipment. The robot should determine when it can safely continue, when it should replan, and when human assistance is required. 

• Investigate learning-based local navigation and decision-making, including reinforcement learning where appropriate, to enable robust behaviour in dynamic and partially predictable environments while respecting explicit safety and operational constraints. 

• Investigate semantic reasoning for difficult recovery situations, using vision-language models as a supervisory layer to interpret the scene and provide contextual information that conventional planners cannot easily capture. For example, the system could distinguish between a temporary obstacle, equipment under maintenance, and a human who should not be interrupted. 

• Integrate the resulting methods into a working robotic system and evaluate them first in simulation and AR-based environments and subsequently on a physical cleanroom mock-up. The experiments will systematically compare planning and recovery strategies using metrics such as transport success rate, recovery time, replanning latency, number of human interventions and safety/constraint violations. 

• The PhD will therefore combine robotics, task and motion planning, autonomous navigation, machine learning and semantic scene understanding to develop robots that can not only follow a plan, but also understand when that plan is no longer appropriate and recover safely. 

Expected expertise domain of the candidate 

Master of Science in electrical engineering, mechatronics, engineering technology, computer science, physics or an equivalent field.  

• Python and C++ 

• A hands-on, build-and-measure attitude as this topic expects you to create real world experiments.  

• Knowledge of multi-agent coordination, centralized and decentralized planning, Task and Motion Planning (TAMP), or Reinforcement Learning (RL) is a plus. 

The research group and team 

The PhD is hosted at imec in collaboration with the Vrije Universiteit Brussel (VUB), at the BruBotics robotics research centre in Brussels. imec is a world-leading R&D hub in nanoelectronics and digital technologies; BruBotics is VUB's robotics centre, working on human-robot interaction, physical safety and mechatronic design. You will join a multidisciplinary international team with access to robot arms and mobile platforms, 3D sensing, motion capture, embedded compute and a mechanical workshop for prototype fabrication. The position is based in Brussels, with regular contact to imec in Leuven, and you will publish at the leading robotics journals (T-RO, RA-M, RA-L) and present at the leading conferences (ICRA, IROS).