Harnessing the flexibility of power-to-heat operation in building and district heating to support electricity systems: A review
Published in: Advances in Applied Energy, Volume 22, 2026, page 100272
Benjamin Freischlad, Chia-Ling Yang, Agnes Engelter, Christian Vossel, Florian Reissner, Florian Steinke, Stefan Niessen
This review examines how power-to-heat (P2H) systems like heat pumps leverage thermal storage and building thermal inertia to provide electricity grid flexibility in district and distributed heating. Using a systematic literature selection process, it surveys simulation tools, control strategies and market frameworks alongside real-world case studies. While small-scale and simulated P2H potential is well-proven, critical gaps persist regarding scalable control strategies, large-scale field validation and cross-sectoral planning.
View at publisher's page
RL finetuning of generalist robot policies under inference latency
Submitted to: 2026 Conference on Robot Learning (CoRL)
Brian Zhu⁵⸍¹, Momen Khalil⁵⸍¹, E. H. Harrison⁵⸍², Emanuele Poggi⁵⸍¹, Philipp Sebastian Schmitt¹, Bernd Kast¹, Philine Meister¹, Pranav Atreya², Qiyang Li², Finn Ferchau¹, Cesar Colmenero¹, Yash Shahapurkar¹, Gokul Narayanan¹, Melih Erdogan¹, Oier Mees³⸍⁴⸍², Kai M. Wurm¹, Georg von Wichert¹, Eugen Solowjow¹, Andrew Wagenmaker², Sergey Levine²
(¹ Siemens, ² UC Berkeley, ³ Microsoft, ⁴ ETH Zurich, ⁵ Equal contribution)
Large modern robot policies, such as vision-language-action (VLA) models, suffer from high inference latency. This delay alters environment dynamics and breaks the Markov assumption, causing standard reinforcement learning (RL) algorithms to fail. To solve this, the authors introduce asynchronous RL with intermediate information (ARLI), a latency-aware framework that enables effective RL fine-tuning despite execution delays. ARLI restores near-Markovian structure using state augmentations that incorporate committed actions and mid-inference observations. Across simulated and real-world tasks, ARLI enables successful RL under latency where standard methods fail, even matching or exceeding ideal zero-latency performance.
View at publisher's page
Co-optimization of inverter controls and line protection functions for improved protection reliability and system stability
Published in: 2025 IEEE Power & Energy Society General Meeting (PESGM)
S. Bhela, S. Gumussoy, U. Muenz (Siemens Foundational Technology, Princeton, NJ, USA), D. J. Kelly, M. J. Reno (Sandia National Laboratories, Albuquerque, NM, USA)
Modern power system protection relies on synchronous generator dynamics, which are increasingly replaced by grid-following and grid-forming (GFM) inverter-based resources (IBRs). Because IBRs display drastically different fault dynamics, existing protection schemes can become inadequate. This paper presents a framework using Bayesian optimization and EMT simulation to co-optimize GFM fault ride-through functions and line protection, demonstrating improved protection reliability and stability on a 100% IBR test network.
View at publisher's page
Smoothed distance functions for direct optimal control of contact-rich systems
Published in: 2025 23rd European Control Conference (ECC)
Christian Dietz, Sebastian Albrecht, Armin Nurkanovic, Moritz Diehl
Standard distance functions like the Euclidean SDF are nondifferentiable for non-smooth rigid bodies like polytopes. This paper proposes a smoothing method for a polytope growth distance function modeled as a parametric linear program. By perturbing optimality conditions, the authors prove the resulting smooth SDF is well-defined. Unlike prior approaches that embed exact SDF optimality conditions directly into an optimal control problem (OCP), smoothing permits external evaluation. This reduces OCP size and distributes computational effort more evenly across function evaluations and solvers. Evaluated via a CasADi implementation on a planar peg-in-hole task, the method demonstrates improved optimization performance.
View at publisher's page
Multi-layer feature exchange transformer for multi-view 6D object pose estimation in robot bin picking
Published in: 2025 IEEE International Conference on Robotics and Automation (ICRA)
Momen Khalil; Vincent Dietrich; Slobodan Ilic
Accurate 6D object pose estimation is vital for robotic bin picking involving textureless, reflective, or occluded items. While multi-view methods outperform single-view approaches, current techniques rely on late-stage fusion, missing full multi-view synergy. This paper introduces a Feature Exchange Transformer (FET) that enables early-stage feature fusion using self-attention and epipolar cross-attention. A coarse-to-fine strategy further optimizes multi-layer feature aggregation across views. Built on top of EpiSurfEmb, the method significantly enhances accuracy and robustness in complex bin-picking scenarios, outperforming current state-of-the-art multi-view methods on the ROBI dataset.
View at publisher's page
