Publications

Manifold Sampling via Entropy Maximization

Cornelius V. Braun*, Tilman Burghoff*, Marc Toussaint; 2026

tl;dr: Sampling from disconnected manifolds is hard; we tackle it with an entropy-maximizing resampling strategy called MASEM.

Under Review

arXiv · Blog · Bibtex

A preview image for the paper. It shows paths that avoid obstacles, sampled via non-linear hit and run (a MCMC sampler). The paths sampled by NHR are visibly less diverse then the paths generated by combining NHR with the method introduced in the paper, MASEM.

Stability-Guided Exploration for Diverse Motion Generation

Eckart Cobo-Briesewitz, Tilman Burghoff, Denis Shcherba, Armand Jordana, Marc Toussaint; 2026

tl;dr: The method (StaGE) generates diverse long-horizon manipulations by leveraging black-box simulation.

Under Review, Presented at ICRA Xplore Workshop 2026 (Oral)

arXiv · Code · Video · Bibtex

The image shows the visual abstract of the paper. It visualizes how StaGE (the method) generates interaction. A large blue sheet shows the manifold of stable configurations, while lines spawing from one point, leaving the manifold, only to end up on it in the end symbolize how StaGE works: The interactions start and end in stability, but are allowed to move through unstable states in between.

Combined Constrained Sampling and Reinforcement Learning for Robotic Manipulation

Marc Toussaint, Cornelius V. Braun, Armand Jordana, Sayantan Auddy, Eckart Cobo-Briesewitz, Denis Shcherba, Tilman Burghoff, Justin Carpentier; 2026

tl;dr: CSRL uses constrained sampling to guide reinforcement learning towards diverse manipulation strategies.

Under Review

arXiv · Website · Video · Bibtex

A preview image for the paper. It shows a robotic hand (the allegro hand) grasping a cube. This image was created during the rollout of a policy trained with the algorithm introduced in the paper.

* Equal Contribution