Massimiliano Carli
Planner-Decoder Separation for Zero-Shot Generalization in Diffusion Policies.
Rel. Giuseppe Bruno Averta, Davide Buoso. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
|
PDF (Tesi_di_laurea)
- Tesi
Accesso limitato a: Solo utenti staff fino al 2 Ottobre 2029 (data di embargo). Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (14MB) |
Abstract
Current goal-conditioned diffusion policies entangle spatial and geometric reasoning with motion execution. A policy trained exclusively on one task memorizes a rigid motor trajectory (e.g. stacking a cube on top of another) and fails on a geometrically related, but topologically distinct, task (e.g., placing the cube to left/right of the other). We propose a modular architecture that decouples planning from action diffusion: a Planner, pretrained via goal-conditioned trajectories sampled from play demostrations, learns how to produce a general and meaningful latent signal, which is then plugged into a diffusion Decoder exclusively conditioned on such output and proprioception data. We argue that a decoder trained exclusively on narrow tasks, conditioned on the continuous signal produced by the pre-trained planner, will still be able to generate new behavior and solve similar tasks zero-shot..
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
