PhD Position on Agentic and System-1/System-2 Architectures for Robot Autonomy
Research Title
Agentic and System-1/System-2 Architectures for Robot Autonomy
Supervisors
Daniele Gammelli, Giuseppe Averta
Topics
Robotics, Machine Learning, Foundation Models, Vision-Language-Action Models, World Models
Funded By
The Italian Institute of Artificial Intelligence (AI4I), in collaboration with Politecnico di Torino
Group Website
https://mira.rd-labs.ai4i.it/
https://danielegammelli.github.io
https://vandal.polito.it
giuseppeaverta.me
Summary Of The PhD Project
The Italian Institute of Artificial Intelligence for Industry (AI4I), in collaboration with Politecnico di Torino, invites applications for a PhD position focused on Agentic and System-1/System-2 Architectures for Robot Autonomy.
This research project investigates how agentic architectures built on multimodal foundation models, Vision-Language-Action (VLA) policies, world models, and model-based controllers can be designed and orchestrated to achieve reliable robot autonomy in unstructured environments. Research will focus on both standalone agentic policies and on System-1/System-2 architectures, in which a slow, deliberative reasoning module (System 2) plans, monitors, and recovers from failures, while a fast, reactive policy (System 1) executes motor commands under real-time and safety constraints. The project focuses on how reasoning, planning, and control should be distributed across these levels, how they should communicate, and how the resulting systems can improve from deployment experience.
Research Objectives And Methods
Recent advances in multimodal foundation models, VLA models, and learned world models are reshaping the design of autonomous robots. Emerging architectures increasingly move beyond fixed perception–planning–control pipelines,
combining large-scale models with specialized policies, tools, memory, and model-based components. However, foundation models are largely trained on data and objectives that do not capture multimodal feedback, real-time constraints, embodiment-specific dynamics, and safety requirements that are central to robotics.
This project studies agentic robotic architectures that exploit the flexibility of foundation models while explicitly accounting for physical constraints, uncertainty, and the limitations of learned components. A natural instantiation is the System-1/System-2 paradigm: a high-level vision-language reasoner decomposes tasks, tracks progress, and decides when and how to intervene, while a low-level VLA or learned skill policy, possibly coupled with optimal control or model predictive control, produces fast and physically consistent actions. Key open questions concern the interface between the two systems (e.g., language, latent plans, subgoals, or trajectories), their different time scales, when the slow system should be invoked, and how failures detected at one level should be handled at the other.
The developed methods will be studied in manipulation and humanoid robotics, with possible extensions to other autonomy domains, and evaluated both in simulation and on real robotic platforms.
Research Objectives
- characterize the capabilities and failure modes of agentic systems for robotics, with emphasis on physical reasoning, long-horizon planning, task completion assessment, failure recovery, and real-time constraints.
- design System-1/System-2 architectures that couple deliberative reasoning models with fast VLA policies and learned skills, studying their interfaces, invocation strategies, and time-scale separation.
- investigate hybrid approaches combining learned components with world models, trajectory optimization, model predictive control, and explicit safety mechanisms.
- study how robotic systems can improve from deployment experience through continual learning, targeted data collection, and data flywheels triggered by failures or novel situations.
- evaluate in simulation and on real robotic platforms, focusing on long-horizon task success, robustness, failure recovery, and data efficiency.
Work Plan Phases
First, the candidate will review the literature on agentic AI, VLA models, world models, and dual-process (System-1/System-2) architectures for embodied AI, and will systematically benchmark current systems to characterize their capabilities and failure modes in physical tasks.
Second, the candidate will develop System-1/System-2 architectures in which a multimodal reasoning model orchestrates one or more VLA policies, learned skills, and model-based controllers. This phase will focus on the representation exchanged between levels, on deciding when deliberation is needed, and on keeping low-level execution reactive and safe.
Third, the candidate will investigate verification and failure-recovery mechanisms across
Agentic and System-1/System-2 Architectures for Robot Autonomy
Supervisors
Daniele Gammelli, Giuseppe Averta
Topics
Robotics, Machine Learning, Foundation Models, Vision-Language-Action Models, World Models
Funded By
The Italian Institute of Artificial Intelligence (AI4I), in collaboration with Politecnico di Torino
Group Website
https://mira.rd-labs.ai4i.it/
https://danielegammelli.github.io
https://vandal.polito.it
giuseppeaverta.me
Summary Of The PhD Project
The Italian Institute of Artificial Intelligence for Industry (AI4I), in collaboration with Politecnico di Torino, invites applications for a PhD position focused on Agentic and System-1/System-2 Architectures for Robot Autonomy.
This research project investigates how agentic architectures built on multimodal foundation models, Vision-Language-Action (VLA) policies, world models, and model-based controllers can be designed and orchestrated to achieve reliable robot autonomy in unstructured environments. Research will focus on both standalone agentic policies and on System-1/System-2 architectures, in which a slow, deliberative reasoning module (System 2) plans, monitors, and recovers from failures, while a fast, reactive policy (System 1) executes motor commands under real-time and safety constraints. The project focuses on how reasoning, planning, and control should be distributed across these levels, how they should communicate, and how the resulting systems can improve from deployment experience.
Research Objectives And Methods
Recent advances in multimodal foundation models, VLA models, and learned world models are reshaping the design of autonomous robots. Emerging architectures increasingly move beyond fixed perception–planning–control pipelines,
combining large-scale models with specialized policies, tools, memory, and model-based components. However, foundation models are largely trained on data and objectives that do not capture multimodal feedback, real-time constraints, embodiment-specific dynamics, and safety requirements that are central to robotics.
This project studies agentic robotic architectures that exploit the flexibility of foundation models while explicitly accounting for physical constraints, uncertainty, and the limitations of learned components. A natural instantiation is the System-1/System-2 paradigm: a high-level vision-language reasoner decomposes tasks, tracks progress, and decides when and how to intervene, while a low-level VLA or learned skill policy, possibly coupled with optimal control or model predictive control, produces fast and physically consistent actions. Key open questions concern the interface between the two systems (e.g., language, latent plans, subgoals, or trajectories), their different time scales, when the slow system should be invoked, and how failures detected at one level should be handled at the other.
The developed methods will be studied in manipulation and humanoid robotics, with possible extensions to other autonomy domains, and evaluated both in simulation and on real robotic platforms.
Research Objectives
- characterize the capabilities and failure modes of agentic systems for robotics, with emphasis on physical reasoning, long-horizon planning, task completion assessment, failure recovery, and real-time constraints.
- design System-1/System-2 architectures that couple deliberative reasoning models with fast VLA policies and learned skills, studying their interfaces, invocation strategies, and time-scale separation.
- investigate hybrid approaches combining learned components with world models, trajectory optimization, model predictive control, and explicit safety mechanisms.
- study how robotic systems can improve from deployment experience through continual learning, targeted data collection, and data flywheels triggered by failures or novel situations.
- evaluate in simulation and on real robotic platforms, focusing on long-horizon task success, robustness, failure recovery, and data efficiency.
Work Plan Phases
First, the candidate will review the literature on agentic AI, VLA models, world models, and dual-process (System-1/System-2) architectures for embodied AI, and will systematically benchmark current systems to characterize their capabilities and failure modes in physical tasks.
Second, the candidate will develop System-1/System-2 architectures in which a multimodal reasoning model orchestrates one or more VLA policies, learned skills, and model-based controllers. This phase will focus on the representation exchanged between levels, on deciding when deliberation is needed, and on keeping low-level execution reactive and safe.
Third, the candidate will investigate verification and failure-recovery mechanisms across