LiftMeUp: Globally optimal algorithms for dexterous manipulation and locomotion

Informacje na temat projektu

LiftMeUp

Identyfikator umowy o grant: 101207106

DOI

10.3030/101207106

Realizacja projektu zakończyła się 31 Grudnia 2025

Data podpisania przez KE 6 Marca 2025

Data rozpoczęcia 1 Maja 2025

Data zakończenia 30 Kwietnia 2027

Finansowanie w ramach

• Marie Skłodowska-Curie Actions (MSCA)

Koszt całkowity

Brak danych

Wkład UE

€ 226 420,56

Koordynowany przez

INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET AUTOMATIQUE
France

Periodic Reporting for period 1 - LiftMeUp (LiftMeUp: Globally optimal algorithms for dexterous manipulation and locomotion)

Okres sprawozdawczy: 2025-05-01 do 2027-04-30

Podsumowanie kontekstu i ogólnych celów projektu

The goal of LiftMeUp was to develop new algorithms to make robots safer, more efficient, and more trustworthy when performing demanding tasks such as dexterous manipulation and reliable locomotion; capabilities that underpin robots’ potential to operate beyond human limits in domains that require scalable, dependable automation. Today’s mainstream options force a trade-off: first-principles, physics-based methods can be transparent but often require extensive tuning and manual heuristics to perform well, while deep-learning approaches are typically data-hungry, less interpretable, and can generalise poorly.
Within the strategic context of the EU’s digital transition and increasing emphasis on AI that is reliable and explainable, the project’s overall objective was to create an easy-to-use framework that uniquely combines data-driven modelling with globally optimal (certifiable) solvers. It aimed to deliver methods that are transparent and sample-efficient, and that reduce sensitivity to initialisation compared with local solvers, which is important for robust performance, safety assurances, and energy/time efficiency.
Methodologically, LiftMeUp progressed in three stages: (1) integrating concepts from Koopman theory, polynomial optimisation, and kernel methods to learn “lifting” functions from data and embed them into globally optimal estimation and control; (2) optimally combining different models into more versatile and expressive solutions that can be updated online; and (3) implementing these algorithms on hardware for real-world locomotion and manipulation tasks. The expected impact is both scientific (new links between machine learning and global optimisation) and practical: more dependable robotics that can be deployed with greater confidence, lower data requirements, and improved resource efficiency, supporting broader uptake of advanced robotics in Europe.

Prace wykonane od początku projektu do końca okresu sprawozdawczego oraz najważniejsze dotychczasowe rezultaty

The project LiftMeUp was terminated early after 8 months, at which point the researcher began a new position as an assistant professor. During the 8-month period, the core technical work to couple data-driven modeling with certifiable (globally optimal) optimization for robotics was conducted. Structured learning pipelines were implemented and tested to identify kernel-based representations amenable to global optimization. The learned kernel models were embedded into globally optimal estimation and control formulations and validated in simulation using standard benchmarks, including inverted-pendulum control and range-only localization. The work produced three peer-reviewed papers, in IEEE T-RO and at the American Control Conference. Building on these foundations, the project initiated application to more challenging contact-rich, dexterous manipulation problems, with publications targeted for 2026.
Another main outcome was a strengthened solver/tooling ecosystem for certifiable optimization: all publications were accompanied by open-source software packages, enabling reproducibility and reuse. The package ksos_tools contains a rank-exploiting SDP solver tailored to sampling-based SDP formulations, and a new toolbox (popcor) was developed to generate SDP relaxations directly from feasible samples.

Innowacyjność oraz oczekiwany potencjalny wpływ (w tym dotychczasowe znaczenie społeczno-gospodarcze i szersze implikacje społeczne projektu)

LiftMeUp delivered core scientific results at the intersection of data-driven modeling and certifiable (globally optimal) optimization for robotics. Key outputs include (i) implemented and validated pipelines for using structured kernel representations for global optimization in robotics; (ii) integration of these models in globally optimal estimation and control, demonstrated on benchmarks such as inverted-pendulum control and range-only localization; and (iii) a strengthened tooling stack for sample-based SDP/SOS optimization, including the reusable software package ksos_tools, with rank-exploiting SDP solver, and the popcor toolbox that generates SDP relaxations from feasible samples. Early steps toward applying these methods in more challenging contact-rich robotics settings were initiated.

The expected impact is to reduce a long-standing bottleneck in robot autonomy: the trade-off between performance (often achieved via black-box learning) and reliability/guarantees (often requiring heavy modeling and manual tuning). By enabling optimization-based estimation/control with stronger correctness guarantees while remaining data-efficient and light on implementation efforts, the project contributes to safer, more predictable behavior in manipulation and locomotion, which is important for higher-confidence deployment in industrial and service robotics.

To ensure further uptake and success, several needs are identified: (1) continued research and engineering to improve scalability of SDP-based methods (solver performance, warm-starting, structure exploitation) and to mature proximal/first-order alternatives where appropriate; (2) demonstration at increasing realism, moving from benchmarks to sustained hardware validation in contact-rich tasks; (3) software integration and usability, e.g. completing robust integration into widely used pipelines and providing documentation, examples, and reproducible benchmarks; (4) consideration of IPR and licensing choices to maximize reuse (open-source core with optional commercial pathways); and (5) alignment with standardization and assurance efforts (interfaces, benchmarking practices, and safety/verification expectations) to position certifiable optimization as a practical ingredient in trustworthy robotic systems.

Overview document of the project, its objectives and relationship with existing approaches