Phd Position On Trustworthy Generalist Robot Policies

9 ore fa

Torino, Piemonte, Italia The Italian Institute Of Artificial Intelligence Tempo pieno

Deadline: 16 November **** at 14:00 CET
Research title: Trustworthy Generalist Robot Policies
Supervisors: Daniele Gammelli, Tatiana Tommasi, Raffaello Camoriano
Topics: Robotics, Machine Learning, Foundation Models, Vision-Language-Action Models, World Models
Funding: Italian Institute of Artificial Intelligence (AI4I), in collaboration with Politecnico di Torino
Group and supervisor websites MIRA Laboratory
VANDAL Laboratory
Daniele Gammelli
Tatiana Tommasi
Raffaello Camoriano
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 Trustworthy Generalist Robot Policies. This research will investigate how generalist robotic systems leveraging Foundation Models (FMs) can be equipped with explicit mechanisms to ensure trustworthy autonomy. It will develop methods to support safe, reliable, and efficient operation while enabling continuous adaptation to novel real-world scenarios and maintaining runtime assurance without compromising generalization and flexibility. The research will focus on trustworthy learned representations, safety specifications, uncertainty estimation, runtime monitoring, and model-based intervention. The proposed methods will be developed and evaluated in representative settings, including autonomous navigation, mobile manipulation, and whole-body robotic systems.
Research Objectives and Methods The rapid advancement of robot learning is enabling generalist policies that leverage Foundation Models and large, heterogeneous datasets, to operate across diverse tasks, objects, environments, and robotic embodiments. However, generalization alone does not guarantee trustworthy autonomy: models optimized for task performance may fail to satisfy explicit safety and operational constraints, particularly in scenarios beyond their training distribution. Reliable deployment requires more than successful task execution.
As robotic systems increasingly rely on learned policies, world models, and multimodal reasoning modules, to estimate risks and support decision-making, a key challenge is determining when their predictions, representations, and intermediate reasoning signals can be trusted in safety-critical settings. Beyond predictive accuracy, trustworthy autonomy requires learned components to faithfully capture relevant system dynamics and causal relationships rather than rely on spurious correlations, and to ensure that generated explanations meaningfully reflect the reasoning underlying the actions ultimately executed. These capabilities are essential for trustworthy autonomous operation in real-world environments.
Research Objectives Characterize and diagnose failures of generalist robotic systems. Investigate the reliability of learned components by assessing whether their predictions, representations, plans, and reasoning signals are consistent with observations, system dynamics, and executed actions, particularly in novel and safety-critical scenarios.
Formalize trustworthiness requirements. Develop methods to specify and evaluate safety, robustness, and efficiency requirements, accounting for geometric, dynamical, semantic, and task-dependent constraints.
Develop runtime assurance mechanisms. Design methods to detect unsafe or unreliable behavior and intervene when necessary by rejecting, modifying, or replacing actions, providing formal safety guarantees or measurable risk reduction while preserving policy generalization.
Enable continual improvement of safety and reliability. Investigate how failures, near-misses, and uncertain situations can inform targeted data collection, policy refinement, and improved safety mechanisms, enabling adaptation to new tasks and environments while maintaining safe operation.
The research work plan will be articulated in four main phases.
First , the candidate will review the literature on Foundation Models for robotic systems (VLA, World Models) and architectures for embodied AI and will systematically benchmark current systems to characterize their capabilities and failure modes in physical tasks.
Second , the candidate will study reliability and faithfulness of learned components used for decision making with methods based on counterfactual interventions, consistency tests, uncertainty estimation, representation analysis, and targeted post-training to identify cases in which apparently plausible predictions or explanations fail to provide reliable evidence about system behavior.
Third , the candidate will develop approaches that combine learned representations with known physical structures, analytical models and explicit constraints to enforce trustworthy operation. Possible approaches will combine learned safety signals with predictive models, constrained optimization, reachability analysis, or control barrier functions to reject, mod