(Poster) Diversifying Parallel Ergodic Search: A Signature Kernel Evolution Strategy
Sreevardhan Sirigiri, Christian Hughes, Ian Abraham, Fabio Ramos openreview.net/pdf?id=3XuUn...
TLDR; Introduced the signature kernel for improving ergodic search

(Spotlight) Accelerating Visual-Policy Learning through Parallel Differentiable Simulation
Haoxiang You, Yilang Liu, Ian Abraham arxiv.org/abs/2505.10646
TLDR; Derived a decoupled policy gradient that accelerates visual-policy learning from sim

CS: bit.ly/ACFRCS25
ECE: bit.ly/ACFRECE25
AMME: bit.ly/ACFRAMME25
BME: bit.ly/ACFRBME25

Check out paper in the repost for more info and results!!!

Think of the contact aware fisher information as a sensitivity measure that we can directly optimize wrt control-contact decisions to help robots learn about their environment more effectively!

Taking a Taylor expansion over parameters yields the contact aware Fisher information matrix that quantifies how useful contact mechanics are to understanding physical constants!!

We posed a contact aware max likelihood problem where the goal was to correlate robot control, sensing to physical (contact) constraints.

indeed. Huge blow to academia and industry. What concerns me is the use of the word "including" feels like a blanket term for an excuse to revoke visas arbitrarily.

***clever -- fun typo

Doing some analysis, we find the observation-based gradient information actually made policy updates more challenging. Removing this term led to an interesting connection with student-teacher policy distillation and huge computational gains! Check out @haoxiang-you.bsky.social's post for more info!

We took a good hard look at the analytical policy gradient going from vision to state to dynamics to policy and realize it is separable!! -- one part only has state/dynamics info, the remaining in observation-only info.

The idea was quite simple, but cleaver --- we are very good/fast at trajectory opt with diff sims, why can't the same be true for visual policy learning?

work done with collaborators across the world from NVIDIA/U Sydney Fabio Ramos, Houston Warren, CMU Geordan Gutow, Ananya Rao, Albert Xu, Howie Choset, and University of Edinburgh Darrick Lee

Elena Wittemyer will present Multi-Agent Ergodic Exploration under Smoke-Based, Time-Varying Sensor Visibility Constraints, WeAT10
lnkd.in/e62mR8X4
TLDR; Integrates ergodic search with a smoke-based fluid solver to inform search of visibility constraints (can help with search in forest fires).

Christian Hughes will present Ergodic Trajectory Optimization on Generalized Domains Using Maximum Mean Discrepancy, ThBT19
lnkd.in/eqQRJnwS
TLDR; A novel metric for ergodic trajectory optimization that only requires domain samples (useful for establishing coverage guarantees for dense SLAM).

Dayi Dong will present Ergodic Exploration over Meshable Surfaces, ThBT19
lnkd.in/eh3NKwdh
TLDR; Demonstrates ergodic trajectory optimization over mesh domains via the Laplace-Beltrami operator.

Organizers: Harish Ravichandar (Georgia Tech), Ian Abraham (Yale), Nadia Figueroa (U Penn), Harry Asada (MIT), and Yasemin Bekiroglu (Chalmers Univ)