Claudio Camolese
PhySVLA: Physics-Informed Safety for Vision Language Action based manipulation.
Rel. Giuseppe Bruno Averta. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (38MB) | Preview |
Abstract
Vision–Language–Action (VLA) models provide general purpose robotic policies capable of interpreting natural language instructions and executing diverse manipulation tasks. However, task success does not imply safe execution: because collision avoidance is generally absent from their training objective, VLAs can produce semantically correct actions that collide with nearby objects. Existing methods either inject geometry during training, requiring access to and adaptation of the policy, or modify inference through policy-internal gradients, simplified obstacle models, or predominantly end-effector centred safety representations. These requirements limit their modularity, scene generality, or full-body protection. We introduce PhySVLA, a task-aware, plug-and-play safety framework that augments frozen VLA policies with physics informed potential field based collision avoidance.
At its core, PhySVLA learns a differentiable neural potential field over the robot configuration space
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Ente in cotutela
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
