Modern robotic manipulation no longer fits into a single discipline. It draws on optimization, control theory, simulation, learned world models and computer vision in equal measure — and the researchers who push the field forward increasingly need to move fluently across all of them. That was the premise of Robotic Manipulation: Foundations to Frontiers (RMFF), a three-day summer school held at CIIRC, Czech Technical University in Prague, from 8 to 10 June 2026, and streamed live to a remote audience.
Over three days, six classes and a series of invited talks took participants from the mathematical foundations of motion generation to the frontiers of learning and perception for robots.
Day 1 — From trajectories to contact
The school opened with sample-based trajectory optimization (Armand Jordana), using stochastic exploration and parallel rollouts to generate robust motions for complex systems. The afternoon turned to rigid-body simulation (Ajay Sathya and Justin Carpentier), covering collision detection, contact dynamics and Simple, a new physics engine built on Pinocchio. Majid Khadiv gave an invited talk on task-and-motion planning with zero-order optimization. The day closed with the AGIMUS final event, followed by a poster session and an evening social.
Day 2 — Control and world models
Nicolas Mansard and Ludovic Righetti led an in-depth class on optimal control, from classical formulations and numerical solvers to whole-body motion. Efstratios Gavves followed with robot learning and world models — building predictive models that let robots reason and plan from experience — with both Righetti and Gavves also giving invited talks.
Day 3 — Perception and planning
The final day moved to vision: Vladimír Petrík taught 6D object pose estimation (from CosyPose to FreePose) for robust grasping, and Josef Sivic gave an invited talk on vision for robotics. Florent Lamiraux closed the program with task and motion planning for long-horizon manipulation, combining symbolic reasoning with continuous motion.
A shared effort
RMFF brought together students and researchers around a single idea: that the next generation of manipulation systems will be built where optimization, learning and perception meet. Across three days, the classes, invited talks and informal discussions made that overlap tangible — and the conversations sparked between participants are exactly what such a school is for. The full program is available at rmffschool.github.io, and the lectures remain online on the school’s YouTube channel.
About the organisers
RMFF 2026 was organised at CIIRC, Czech Technical University in Prague, and hosted by the ELLIS Unit Prague. The school was made possible by the following EU-funded research projects.
ELLIOT (GA No. 101214398) develops large-scale, open multimodal foundation models with strong spatio-temporal understanding, leveraging European high-performance computing and combining real and synthetic data to support robust, general-purpose AI across diverse modalities.
AGIMUS (GA No. 101070165) delivers open-source breakthrough innovation in AI-powered agile production, advancing perception, planning and control so that general-purpose robots can be deployed quickly, operate
autonomously and adapt to changing manufacturing processes.
ERC FRONTIER (GA No. 101097822) is an ERC Advanced Grant led by Josef Sivic within the IMPACT group at CIIRC, advancing embodied AI across machine perception, robot learning and AI-driven protein engineering.
ELIAS (GA No. 101120237) is the European Lighthouse of AI for Sustainability, a Network of Excellence connecting leading academic researchers with industry practitioners to advance AI that drives sustainable innovation while supporting a cohesive society and respecting individual rights.
