Constrained Sampling to Guide Universal Manipulation RL

Marc Toussaint 1,2 , Cornelius V. Braun 1 , Eckart Cobo-Briesewitz 1 , Sayantan Auddy 1 , Armand Jordana 3 , Justin Carpentier 4 · 2026

Read Constrained Sampling to Guide Universal Manipulation RL by Marc Toussaint 1,2 , Cornelius V. Braun 1 , Eckart Cobo-Briesewitz 1 , Sayantan Auddy 1 , Armand Jordana 3 , Justin Carpentier 4 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-08557.txt.

Description

Teaching robots to push and slide objects without grasping them is hard, because reinforcement learning struggles to discover contact-rich strategies. Marc Toussaint and seven co-authors combine learning with a smarter way of choosing where episodes start. Their arXiv paper targets non-prehensile manipulation in settings full of contact.

A new state sampler covers diverse contact modes by respecting contact structure, and resets drawn from it are paired with projected interpolation and curriculum learning. The combination beats plain reinforcement learning and alternative reset schemes, training universal, non-prehensile, and dynamic policies. Readers get a practical boost for robot manipulation that leans on contact rather than gripping.

Catalog record

CREATORMarc Toussaint 1,2 , Cornelius V. Braun 1 , Eckart Cobo-Briesewitz 1 , Sayantan Auddy 1 , Armand Jordana 3 , Justin Carpentier 4
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-08557

Browse the full library · About the Tower of Babel