2025Activity report‌​‌Project-TeamHEPHAISTOS

7.1.1 Parametric​​​‌ equations solver with AI​

Participants: Jean-Pierre Merlet.​‌

These last 3 years​​ we have worked on​​​‌ a generic solver for​ finding the real root(s)​‌ of parametric non linear​​ square system of equations​​​‌ $𝐅(𝐗,\Lambda )=\mathbf{0‌}$ i.e. systems which have​​ as many unknowns $\mathrm{𝐗‌}$ as equations but with​ equation coefficients that are​‌ functions of a parameters​​ set $\Lambda$ whose elements​​​‌ are assumed to be​ bounded. Such a system​‌ has usually several solutions​​ and their number depends​​​‌ on the parameters values​ and cannot be determined​‌ in advance. Our aim​​ is to obtain a​​​‌ solver that provides exact​ solutions (i.e. as close​‌ as needed). We also​​ aim at getting all​​​‌ solutions for any $\mathrm{\Lambda }$: we cannot guarantee​‌ this point for our​​ solver but it has​​​‌ a self-learning capacity that​ allows to reduce the​‌ number of missed solutions.​​

Beside the equations there​​​‌ is a single entry​ point for building our​‌ solver: we assume that​​ we have been able​​​‌ to determine an initial​ solution set i.e. the​‌ solutions (not necessarily all​​ of them) for a​​​‌ small set $V_{\mathrm{i}}=\{{\Lambda }_{\mathrm{1‌}},...{\Lambda }_{\mathrm{n}}\}$ of $\Lambda$ of​​​‌ specific systems (typically this​ set has 8 samples​‌ of $\Lambda$). Using​​ continuation and singularity checking,​​​‌ we can split the​ initial solution set into​‌ different components. For each​​ of these components, we​​​‌ may create arbitrary large​ learning sets, i.e.​‌ set of samples (​​${\Lambda }_i,{\mathrm{𝐒‌}}_i$) where ${\mathrm{𝐒}}_i$ is a solution​‌ of the system for​​ ${\Lambda }_i$. We​​​‌ then use these learning​ sets to create Multi-Layer​‌ Perceptrons (MLP) but we​​ use a specific training​​​‌ strategy based on the​ concept of success rate​‌: the success rate​​ is the percentage of​​​‌ samples whose solutions are​ obtained by using the​‌ prediction of the MLP​​ as a guess for​​​‌ a deterministic method (typically​ the Newton method). The​‌ training strategy mixes the​​ decrease of a loss​​​‌ function but the trained​ MLP will be the​‌ one that has the​​ lowest loss but also​​​‌ the largest success rate.​ If this success rate​‌ is 100, then the​​ MLP hybridized with the​​ Newton method will find​​​‌ the solution ${𝐒}_{\mathrm{i‌}}$ for all sample of‌​‌ the learning set. If​​ the success rate is​​​‌ not 100, we use‌ the missed solutions to‌​‌ create a specific learning​​ set for a new​​​‌ MLP. The training stops‌ when all solutions of‌​‌ all learning sets are​​ found when submitted to​​​‌ the created MLPs. Hence‌ the solving process consists‌​‌ in submitting an input​​ ${\Lambda }_j$ to all​​​‌ the hybridized MLPs (but‌ the Newton method is‌​‌ run only if the​​ MLP predictions has some​​​‌ sense). This solver may‌ be improved by creating‌​‌ verification sets$\{{\mathrm{\Lambda }}_j,{𝐒}_{\mathrm{𝐣‌}}^{\mathbf{1}},{𝐒}_{\mathrm{𝐣‌}}^{\mathbf{2}},...{\mathrm{𝐒‌‌}}_{𝐣}^{𝐧}\}$,where​​ the ${𝐒}_{𝐣}^{\mathrm{𝐤‌}}$ are different solutions of‌ the system for $\mathrm{\Lambda ‌‌}={\Lambda }_j$.​​ They are derived from​​​‌ the initial solution set‌ in such a way‌​‌ that the ${\Lambda }_{\mathrm{j}}$ are different from the​​​‌ one in the learning‌ sets. If solution(s) are‌​‌ missed by the solver​​ they are used to​​​‌ create new hybridized MLPs‌ that are added to‌​‌ the solver.

Although creating​​ the solver may require​​​‌ several hours, the computation‌ time for getting the‌​‌ solutions for a given​​ ${\Lambda }_j$ is very​​​‌ low, typically a few‌ ms (furthermore parallel computing‌​‌ allows for a drastic​​ decrease for both training​​​‌ the solver and for‌ getting the solutions). Hence‌​‌ the solver should be​​ used when it is​​​‌ expected to solve the‌ system for a large‌​‌ number of ${\Lambda }_{\mathrm{j}}$.

This solver has​​​‌ initially been developed for‌ solving the direct kinematics‌​‌ of cable-driven parallel robot​​ with sagging cables. Alternative​​​‌ solving methods, such as‌ interval analysis or continuation,‌​‌ require several hours so​​ that with a training​​​‌ time of about 60‌ hours the proposed solver‌​‌ is more efficient as​​ soon as more than​​​‌ 20 occurrences of the‌ problem are considered. The‌​‌ solver has also been​​ tested on difficult problems​​​‌ from the ALIAS benchmark‌ such as the Brent‌​‌ reactor problem with similar​​ success.

We have succeeded​​​‌ in extending the principle‌ of the proposed solver‌​‌ to deal with ODE​​ (that appear, for example,​​​‌ in the kinematics of‌ robot with flexible limbs,‌​‌ see section 7.2.1).​​ We have also shown​​​‌ that the solver may‌ be used for non-square‌​‌ system with $n$ equations​​ and $m>\mathrm{n‌}$ unknowns leading to an‌ infinite number of solutions‌​‌ as soon as optimizing​​ some criterion should lead​​​‌ to a unique solution.‌ Indeed we may develop‌​‌ a gradient-descent method that​​ consider all combinations of​​​‌ $m-n$ unknowns‌ and set a fixed‌​‌ values for them. In​​ that case, the solver​​​‌ is able to provide‌ all possible values for‌​‌ the remaining unknowns so​​ that the optimizer is​​​‌ able to determine the‌ best descent direction. We‌​‌ have tested this approach​​ on the inverse kinematics​​​‌ of cable-driven parallel robot‌ with more than $\mathrm{n‌‌}>6$ sagging cables.​​ In that case, we​​​‌ have $2n+‌6$ equations for $\mathrm{3‌‌}n$ unknowns so that​​​‌ the system is under-constrained​ as soon as $\mathrm{n‌}>6$. The​​ proposed approach leads to​​​‌ a criteria value which​ differs by at most​‌ 2% from the optimum​​ value with a computation​​​‌ time which is at​ least 10 times lower​‌ than any other optimizer.​​ This year we have​​​‌ tested this approach on​ a calibration problem. The​‌ cables in a cable-driven​​ parallel robot are submitted​​​‌ to wear that leads​ to a decrease of​‌ their Young modulus and​​ it is interesting to​​​‌ get an estimation of​ the modulus first for​‌ preventive maintenance but also​​ for improving the positioning​​​‌ accuracy of the platform.​

Being given a platform​‌ pose $𝐗$ and its​​ corresponding cable lengths at​​​‌ rest, these decreases lead​ to a modification of​‌ the effective pose $\widehat{\mathrm{𝐗}}$ that is reached​​​‌ by the platform. We​ assume that the effective​‌ platform pose can be​​ measured at least at​​​‌ some calibration poses${\mathrm{𝐗}}_c^i$ where the​‌ platform is basically supported​​ by 2 or 3​​​‌ cables, called supportive cables​, whose tensions is​‌ much higher than the​​ one of the remaining​​​‌ cables (the measurement assumption​ is valid for our​‌ prototypes that are equipped​​ with lidars on the​​​‌ platform). The $n$ calibration​ poses are chosen in​‌ such a way that​​ all cables are supportive​​​‌ in at least one​ of these poses. We​‌ then define a cost​​ function as ${\sum }_{\mathrm{j‌}=1}^{j=n}\left|\right|{\mathrm{𝐗‌}}_c^j\widehat{{𝐗}_{\mathrm{c}}^j}{||‌}^2$ with the purpose​ of finding the Young​‌ modulus $E_i$ of​​ each cable that minimize​​​‌ the cost function. Using​ classical optimizers and taking​‌ into account that there​​ may be uncertainties on​​​‌ the pose measurements and​ on the cable lengths,​‌ we have been able​​ to get reasonable $\mathrm{E‌}$ estimation as soon as​ the decrease for at​‌ least one of the​​ $E$ was about 10-15%​​​‌ compared to the $\mathrm{E}$ when the cable is​‌ new, but the process​​ is computationally intensive (several​​​‌ hours). Including the direct​ kineematics neural network solver​‌ in the optimization process​​ allows one to get​​​‌ an almost real-time estimation​ of the $E$ as​‌ soon as the decrease​​ is about 20-30% so​​​‌ that the $E$ estimation​ may be used immediately​‌ to improve the positioning​​ accuracy of the robot​​​‌ 18.

Our objective​ now is to develop​‌ a generic framework for​​ obtaining a solver for​​​‌ arbitrary square system of​ equations with minimal manual​‌ intervention of the end-user.​​

7.2.1 Kinematics of​​ soft robots and AI​​​‌

Participants: Jean-Pierre Merlet.​

There is a growing​‌ academic interest in soft​​ robots for which the​​​‌ links are made of​ flexible material, many of​‌ which have a closed-loop​​ structure (typically like parallel​​​‌ robots). A community (that​ includes the INRIA DEFROST​‌ project-team in Lille 20​​) is working on​​​‌ soft robots, having benefited​ from 3D printing to​‌ create numerous prototypes, and​​ focuses on the complex​​ modeling of such a​​​‌ system. From a kinematic‌ viewpoint, the main topic‌​‌ is inverse kinematic solution​​ (finding the actuators positions​​​‌ to have the end-effector‌ of the robot in‌​‌ a given pose). Yet,​​ complex issues computing the​​​‌ robot workspace, finding all‌ solutions of the direct‌​‌ kinematics or solving design​​ problem (e.g. determining the​​​‌ geometry(ies) of the robot‌ so that all poses‌​‌ in a given workspace​​ are reachable) are rarely​​​‌ addressed excepted for planar‌ robots. Basically flexible beams‌​‌ are modeled through an​​ ODE but, as an​​​‌ analytic form for its‌ solution is usually not‌​‌ known, a discrete approximation​​ is used for numerical​​​‌ treatment and consequently the‌ solver proposed in section‌​‌ 7.1.1 may be used.​​ An invitation from O.​​​‌ Petuya and M. Urizar‌ Arana at Bilbao University‌​‌ has allowed us to​​ get familiar with the​​​‌ problem of dealing with‌ a flexible 6 degrees‌​‌ of freedom parallel robot​​ with 6 flexible links​​​‌ with an extremity that‌ is fixed on an‌​‌ actuated revolute joint. Each​​ beam is decomposed into​​​‌ 10 elements and that‌ lead to a relatively‌​‌ large system of equations.​​ However, with about 300​​​‌ hybridized MLPs, we have‌ been able to find‌​‌ all solutions on verification​​ sets with a total​​​‌ of 3000 direct kinematics‌ solutions (for the specific‌​‌ robot we are considering,​​ the direct kinematics may​​​‌ have from 5 to‌ 14 solutions). Regarding the‌​‌ computation time, the sequential​​ version of the solver​​​‌ requires about 30 seconds‌ to provide the direct‌​‌ kinematics solutions. However, a​​ distributed implementation on IA​​​‌ processors with multiple GPU‌ will drastically reduce this‌​‌ time. Our collaborator of​​ University of Bilbao have​​​‌ implemented a direct kinematics‌ solver based on an‌​‌ optimization process but the​​ neural network solver is​​​‌ about 10 times faster‌ and in some cases‌​‌ provides direct kinnematics solutions​​ that have not been​​​‌ found by the Bilbao‌ solver.

Our objective now‌​‌ is to work on​​ a generic framework with​​​‌ a distributed implementation that‌ will be used both‌​‌ for the training of​​ the MLP and for​​​‌ the solver.

7.2.2 Green‌ robotics

Participants: Jean-Pierre Merlet‌​‌, Yves Papegay,​​ Clara Thomas [correspondant].​​​‌

About 48% of the‌ 4 000 000 existing‌​‌ robots are performing manutention​​ operations and about half​​​‌ of them are performing‌ repetitive pick-and-place operations, where‌​‌ an object has to​​ be moved from A​​​‌ to B. This motion‌ usually involves only 3‌​‌ translational degrees of freedom​​ and possibly a rotation​​​‌ around the vertical axis.‌ The total energy consumption‌​‌ of the existing robots​​ was evaluated to be​​​‌ 6 705 GWh in‌ 2022, which is roughly‌​‌ 0.013% of the world​​ energy consumption of industry.​​​‌ It may be conservatively‌ estimated that pick-and-place robots‌​‌ involves about 1500 GWh.​​ Although robotics is not​​​‌ a major player regarding‌ energy consumption, it is‌​‌ still interesting to design​​ robots which use less​​​‌ energy and resources.

Several‌ types of mechanical architecture‌​‌ are used for pick-and-place​​ operation: serial, cartesian and​​​‌ Scara types. These architectures‌ are not energy efficient‌​‌ as they impose to​​​‌ actuate several heavy mechanical​ elements beside the load​‌ (e.g the serial type​​ require energy just to​​​‌ stay in its pose).​ Furthermore, these robots are​‌ controlled with a rather​​ powerful computer as they​​​‌ may be used for​ other purposes and these​‌ systems use a lot​​ of mechanical and electronic​​​‌ resources while being not​ very flexible and difficult​‌ to maintain. All in​​ all, this leads to​​​‌ a relatively large energy​ consumption and an extensive​‌ use of resources for​​ such a simple task.​​​‌ Our objective is to​ propose a specific robot​‌ for pick-and-place operations with​​ a much lower energy​​​‌ consumption, less resources, being​ simple to maintain and​‌ offering a larger flexibility​​ than the currently used​​​‌ robots.

The first step​ of our proposal is​‌ to adopt a different​​ mechanical architecture based on​​​‌ cable-driven parallel robots where​ the robot's end-effector is​‌ connected to the ground​​ by cables whose lengths​​​‌ is controlled. They have​ multiple advantages:

• they involve​‌ only a very limited​​ mobile mass as beside​​​‌ the load only the​ cables are moving
• they​‌ are highly flexible: just​​ by moving the winches,​​​‌ it is possible to​ cover any kind of​‌ workspace even a very​​ large one (the FAST​​​‌ telescope covers a circular​ area of 500 m​‌ of diameter)
• their lifting​​ capacity may be very​​​‌ high (our Marionet-crane has​ a lifting capacity of​‌ 2.5 tons)
• they are​​ easy to maintain either​​​‌ by just changing the​ cables or the winches​‌ that are fairly standard​​
• the energy efficiency or​​​‌ parallel robots is, in​ general, 25% higher than​‌ any other robot structure​​

Cable-driven parallel robots (CDPR)​​​‌ may be designed for​ various degrees of freedom​‌ but the simpler one,​​ called the $N$-1​​​‌ CDPR, has $N$ cables​ attached at the same​‌ point on the platform.​​ This CDPR offers either​​​‌ planar motion with the​ 2-1 ($x-‌z$ motion) or spatial​​ motion with the 3-1​​​‌ and 4-1 (the difference​ being a larger workspace​‌ for the 4-1). The​​ platform may include an​​​‌ independent rotation motion if​ necessary. Our approach excludes​‌ very fast pick-and-place operation​​ that are usually performed​​​‌ with parallel robot of​ the Delta type: CDPR​‌ may also be fast​​ but we are more​​​‌ interested in pick-and-place operation​ over a large workspace​‌ and possibly for heavy​​ load (typically classical pick-and-place​​​‌ robots are designed for​ a load of 1​‌ kg).

The image shows​​ an indoor setup with​​​‌ two tripods positioned in​ front of a gray​‌ background. Each tripod holds​​ a spool of red​​​‌ thread. The threads are​ tensioned and connected together​‌ to a black part​​ to mimick a pick​​​‌ and place operation.

To check the energy​ efficiency of a CDPR,​‌ we have considered a​​ 2-1 CDPR (see Figure​​​‌ 1) and we​ have defined a classical​‌ pick-and-place trajectory (a vertical​​ upward motion from $\mathrm{A‌}$ to $A_v$,​ an horizontal motion toward​‌ a position $B_{\mathrm{v}}$ at the vertical from​​ B and a vertical​​​‌ downward motion toward B).‌ In parallel we have‌​‌ collected a database of​​ energy consumption for pick-and-place​​​‌ robots and its partition‌ between the robot and‌​‌ the control cabinet, the​​ later one representing in​​​‌ general around 40% of‌ the total energy consumption.‌​‌ Note that most of​​ these robots have usually​​​‌ a relatively small workspace‌ (the distance between $\mathrm{A‌‌}$ and $B$ being less​​ than one meter). Last​​​‌ year, we have exhibited‌ a theoretical model of‌​‌ energy consumption for the​​ 2-1, 3-1 and 4-1​​​‌ CDPR. This model has‌ been experimentally validated for‌​‌ the 2-1 CDPR, showing​​ that the robot's energy​​​‌ consumption was about 500‌ times lower than classical‌​‌ robots for the same​​ payload.

Regarding energy consumption,​​​‌ we have still used‌ a computer to control‌​‌ our CDPR although a​​ low power computer may​​​‌ be used (a Raspberry‌ Pi is largely sufficient).‌​‌ Regarding electronic resources, we​​ were using Phidgets boards​​​‌ (one for controlling the‌ velocities of the two‌​‌ motors and one for​​ getting the values of​​​‌ the motor's encoders). For‌ grasping the object at‌​‌ the pick place, we​​ have on the platform​​​‌ a battery and a‌ radio receiver with a‌​‌ relay that switches on​​ an on-board electric magnet:​​​‌ this relay is activated‌ by the computer which‌​‌ has a radio emitter.​​ Our objective this year​​​‌ was to get rid‌ of the computer and‌​‌ of most of the​​ electronic resources.

For that​​​‌ purpose we have calculated‌ the cable velocities $\dot{\mathrm{\rho ‌‌}}(t)$ as function of time​​​‌ $t$ to execute the‌ pick-and-place trajectory at a‌​‌ given speed between $\mathrm{A}$ and $B$. We​​​‌ have then designed and‌ 3D-printed a cylindrical rotating‌​‌ drum, called the control​​ drum (displayed in Figure​​​‌ 2), with two‌ tracks (one for each‌​‌ motor) that are followed​​ by resistive sliders.

The​​​‌ image shows a mechanical‌ device on a table.‌​‌ Attached to a metallic​​ frame is a red​​​‌ circular drum connected to‌ two large gears connected‌​‌ by a belt to​​ a motor. Several sensors​​​‌ acts on an engravure‌ of the red drum.‌​‌

The sliders motions reflect‌​‌ exactly the function $\dot{\mathrm{\rho }}(t)‌$ to get from $\mathrm{A‌}$ to $B$ and then‌​‌ from $B$ to $\mathrm{A}$ to perform a full​​​‌ cycle so that we‌ get from the sliders‌​‌ two control tensions that​​ are sent to a​​​‌ motor controller board. Being‌ given the clearance of‌​‌ the sliders, the platform​​ motion is approximate and​​​‌ the positioning accuracy degrades‌ when multiple cycles are‌​‌ performed. Hence we must​​ design a system to​​​‌ ensure that the motion‌ to $A$ and $\mathrm{B‌‌}$ is accurate. For that​​ purpose, we use 4​​​‌ switches on the robot‌ drums that is moving‌​‌ when the cable lengths​​ are changed. Two of​​​‌ them $S_1^{\mathrm{j‌}},S_2^{\mathrm{j‌‌}}$ for cable $j$ are​​ used to detect that​​​‌ the $j$-th cable‌ length corresponds to the‌​‌ one for having the​​​‌ platform close to ${\mathrm{A}}_v,B_{\mathrm{v‌}}$ and the other two,​​ $S_3^j,‌S_4^j$ play​ the same role for​‌ $A$ and $B$.​​ Hence we have the​​​‌ following triplets (platform pose,​ switch activated on cable​‌ 1, switch activated on​​ cable 2): $(\mathrm{A‌},S_3^{\mathrm{1}},S_3^{\mathrm{2‌}})$, $(\mathrm{B},S_4^{\mathrm{1‌}},S_4^{\mathrm{2}})$, $({\mathrm{A‌}}_v,S_{\mathrm{1}}^1,S_{\mathrm{1‌}}^2)$, (​$B_v,{\mathrm{S‌}}_2^1,{\mathrm{S}}_2^2$). Starting​​​‌ from $A$, we​ bypass the tensions provided​‌ by the sliders and​​ send to the motor​​​‌ controller two constant tensions​ that allow to get​‌ an approximate vertical motion​​ of the platform: 0​​​‌ to motor $k$ as​ soon as switch ${\mathrm{S‌}}_1^k$ is activated.​​ When both switch ${\mathrm{S‌}}_1^1,{\mathrm{S}}_1^2$ are active​‌ (the platform is at​​ $A_v$), we​​​‌ switch back to the​ slider tensions. As soon​‌ as $S_2^{\mathrm{k}}$ is activated, we stop​​​‌ motor $k$ and when​ both $S_2^{\mathrm{1‌}},S_2^{\mathrm{2}}$ are active (the platform​​​‌ is at $B_{\mathrm{v}}$) we bypass the​‌ slider tensions and send​​ a constant tension to​​​‌ each motors that are​ stopped whenever $S_{\mathrm{4‌}}^1$ or $S_{\mathrm{4}}^2$ are active so​​​‌ that the platform stops​ at $B$. This​‌ process is reverted to​​ move from $B$ to​​​‌ $A$.

In term​ of energy consumption, the​‌ system uses only energy​​ to move the load​​​‌ along the trajectory and​ to rotate continuously the​‌ control drum and needs​​ only the motor controller​​​‌ board as electronic resource.​ It is still programmable:​‌ the trajectory speed may​​ be adjusted by changing​​​‌ the rotation speed of​ the control drum and​‌ executing a new trajectory​​ require only printing a​​​‌ new control drum (we​ have designed a software​‌ that takes as input​​ the trajectory and generates​​​‌ printing data). The system​ is also easy to​‌ maintain and repair with​​ only 3 motors, one​​​‌ electronic board and a​ few mechanical elements (the​‌ cables and the reduction​​ gears for the motors).​​​‌ It can also accommodate​ large loads by changing​‌ the power of the​​ two motors actuating the​​​‌ cable drums.

The system​ has been tested for​‌ two trajectories respectively with​​ a length of 80​​​‌ cm and 400 cm​ (the later being difficult​‌ to perform with classical​​ robots) and a load​​​‌ of 1 kg: the​ energy consumption is 100​‌ lower than the one​​ of the best classical​​​‌ robot. The main issue​ of the 2-1 is​‌ oscillation of the load,​​ which is however small​​​‌ at the pick and​ place pose. A simple​‌ way to reduce this​​ oscillation is to have​​​‌ 2 cables on each​ motor drum that act​‌ as a parallelogram with​​ a drastic reduction of​​​‌ the oscillation. For next​ year, we plan to​‌ test a 3-1 CDPR​​ that has a spatial​​ workspace allowing to avoid​​​‌ obstacles between $A$ and‌ $B$.

Note that‌​‌ we may also consider​​ the case where the​​​‌ sliders of the control‌ drum may directly actuate‌​‌ the winches. This system​​ will use single motor​​​‌ and no electronic board‌ while a simple computer‌​‌ is used to design​​ the control drum: we​​​‌ will be back to‌ a mechanically programmable robot‌​‌ !

7.2.3 Multi-room assistive​​ mobility robot

Participants: Jean-Pierre​​​‌ Merlet, Yves Papegay‌, Clara Thomas [correspondant]‌​‌.

The project-team has​​ a long experience on​​​‌ using cable-driven parallel robots‌ (CDPR) for assisting mobility‌​‌ of frail people. We​​ have already shown, both​​​‌ theoretically and experimentally, that‌ this type of robot‌​‌ has several advantages for​​ this purpose: modularity (adapting​​​‌ the robot geometry to‌ manage both the end-user‌​‌ and its environment), large​​ lifting capacity, low intrusivity​​​‌ and cost. Such a‌ robot is in concurrence‌​‌ with other assistive robots,​​ such as a Cartesian​​​‌ robot with rails on‌ the ceiling (very intrusive‌​‌ and high cost) or​​ mobile robots (very intrusive,​​​‌ low energy autonomy, limited‌ mobility and not appropriate‌​‌ for moving the subject​​ in small rooms like​​​‌ in toilets or bathrooms).‌ But a common drawback‌​‌ of the CDPR or​​ Cartesian robots is that​​​‌ they are limited to‌ be used in a‌​‌ single room and it​​ is not reasonable to​​​‌ have one such robot‌ in every room (with‌​‌ the problem for the​​ end-user to use these​​​‌ robots in sequence for‌ example when the subject‌​‌ is in a harness).​​

Hence one of our​​​‌ objective for the coming‌ years is to develop‌​‌ a 4-1 CDPR (which​​ allows the end-user to​​​‌ move in any part‌ of a rectangular room)‌​‌ whose geometry (i.e. the​​ location of the pulling​​​‌ points of the cables)‌ may be changed on‌​‌ the fly to realize​​ a strategy allowing the​​​‌ end-user to move from‌ one room to another‌​‌ one, the rooms being​​ joined by a corridor.​​​‌

Changing a CDPR geometry‌ -

There are two‌​‌ options to change the​​ geometry of a CDPR:​​​‌ moving the winches or‌ using pulleys to redirect‌​‌ the cables toward their​​ pulling points. For moving​​​‌ a winch, we may‌ have it on a‌​‌ wheeled mobile platform (supported​​ by 2 parallel rails)​​​‌ that has its own‌ battery and will stop‌​‌ at specific places where​​ electrical contacts allows one​​​‌ to supply power for‌ the winch (see Figure‌​‌ 3). For changing​​ the direction of the​​​‌ mobile platform (e.g. at‌ a corner), we use‌​‌ swing bridges with a​​ linear actuator to control​​​‌ the platform rotation.

The‌ image shows a mechanical‌​‌ system as a whole​​ and a close-up. It​​​‌ consist of a moving‌ platform based on wheels‌​‌ on a double rail​​ that is able to​​​‌ turn to ensure connection‌ between two perpendicular linear‌​‌ segments.

The image shows​​ a mechanical system as​​​‌ a whole and a‌ close-up. It consist of‌​‌ a moving platform based​​ on wheels on a​​​‌ double rail that is‌ able to turn to‌​‌ ensure connection between two​​​‌ perpendicular linear segments.

In the pulleys version,​‌ the 4 winches are​​ fixed on the ground​​​‌ and the cable follows​ a circuit determined by​‌ the position of the​​ pulleys which are either​​​‌ in a fixed position​ (e.g. to round a​‌ corner) or are mobile​​ (for the pulley leading​​​‌ to the outpoint point​ of the cable). To​‌ avoid intrusivity, the cables​​ follow the walls until​​​‌ they reach an output​ point. Note that for​‌ both versions, a cable​​ output point may be​​​‌ moved only if the​ cable tension is very​‌ low.

7.2.4 Networks of‌​‌ conversational agents and robots​​

Participants: Yves Papegay.​​​‌

The use of Large‌ Language Models (LLMs) in‌​‌ robotics is an emerging​​ trend. These new natural​​​‌ language processing capabilities aim‌ to improve the induction‌​‌ of high-level robot behavior.​​ The proposed approach relies​​​‌ on conversational agents organized‌ in graphs, a powerful,‌​‌ modular, and flexible method​​ that leverages LLMs to​​​‌ interpret and contextualize commands‌ while integrating classical components‌​‌ for motion planning, dynamic​​ environment control, or object​​​‌ manipulation. The first challenge‌ is to reliably transform‌​‌ descriptions of the environment,​​ instructions, or constraints expressed​​​‌ in natural language into‌ a functional decomposition of‌​‌ robotic tasks. This requires​​ combining the contextualization provided​​​‌ by LLM embedding mechanisms‌ with access to factual‌​‌ information about the environment,​​ whether dynamic (from sensors)​​​‌ or static. A second‌ challenge lies in the‌​‌ increasing complexity of agent​​ organization, which grows with​​​‌ the complexity of the‌ tasks to be managed.‌​‌ This organization must dynamically​​ adapt to changes in​​​‌ the robots behavior or‌ capabilities while ensuring the‌​‌ correct execution of tasks.​​ With specialized agents possessing​​​‌ domain expertise, these technologies‌ promise to enhance robots'‌​‌ ability to adapt to​​ complex situations and provide​​​‌ robust solutions to the‌ application problems encountered.

This‌​‌ new exploration work started​​ last year in collaboration​​​‌ with David Daney (Auctus‌ team, centre Inria de‌​‌ l'Université de Bordeaux). It​​ was supposed to take​​​‌ place and to be‌ supported by a robotics‌​‌ PEPR proposition that has​​ been delayed for one​​​‌ year.

7.3.1 Monte-Carlo‌​‌ simulation models

7.4.1 Apeiron: From Biofilm‌​‌ Eradication to the Emergence​​ of Artificial Life

The​​​‌ artwork Apeiron was created‌ in response to a‌​‌ private commission from the​​ owner of the La​​​‌ Salle Blanche (LSB) factory‌ in Apt, which designs‌​‌ and manufactures decontaminable furniture​​ for clean rooms (laboratories​​​‌ and operating theaters). To‌ protect the south-facing façade‌​‌ of the workshops from​​ direct sunlight, the commission​​​‌ consists in installing a‌ micro-perforated tarpaulin measuring 42‌​‌ meters in length and​​ 4.6 meters in height,​​​‌ printed each year with‌ a different artistic proposal.‌​‌ In 2025, Apeiron inaugurates​​ this ephemeral collection.

The​​​‌ image shows the facade‌ of the fatory with‌​‌ the printed tarpaulin of​​ the artwork.

While LSB’s​​ furniture is designed to​​​‌ delay the appearance of‌ biofilms and to facilitate‌​‌ their destruction on white​​ surfaces, our artistic- scientific​​​‌ collaboration took an opposite‌ approach, focusing on the‌​‌ spontaneous emergence of micro-cellular​​ behaviors that model living​​​‌ systems through continuous cellular‌ automata of the Lenia‌​‌ type.

At the intersection​​ between destruction and the​​​‌ spontaneous creation of living‌ cellular behaviors, the artwork‌​‌ consisted in designing a​​ cellular automaton scaled to​​​‌ the dimensions of the‌ printable surface and following‌​‌ the color codes of​​​‌ biofilm images obtained by​ fluorescence microscopy, notably by​‌ the Institut Pasteur: inert​​ surfaces in gray, living​​​‌ cells in fluorescent green,​ and dead cells in​‌ red (see Figure 4​​).

8.1.1​​ Other european programs/initiatives

• Hephaistos​​​‌ is part of the​ euROBIN, the Network of​‌ Excellence on AI and​​ robotics that was launched​​​‌ in 2021.

9.1.1 Scientific events:​ organization

9.1.2 Journal

9.1.3 Invited talks

• Jean-Pierre​ Merlet has been invited​‌ for one week at​​ Bilbao University for presenting​​​‌ his work on neural​ networks and kinematics and​‌ to collaborate for its​​ use for continuum robot​​​‌ (section 7.2.1).
• Jean-Pierre​ Merlet has been invited​‌ to the workshop Regards​​ Croisés en SHS sur​​​‌ les Exosquelettes organized at​ Artois University, both to​‌ give a talk on​​ robotics assistance device and​​​‌ to provide technical comments​ on the works of​‌ the human science community​​ on the topic of​​​‌ exoskeletons.
• Jean-Pierre Merlet has​ been invited as a​‌ panelist during the conference​​ CableCon in Hong-Kong and​​​‌ gave an invited talk​ during the workshop MOMI​‌ on Human Computer Interaction.​​

9.1.4 Leadership within the​​​‌ scientific community

• Jean-Pierre Merlet​ is a member of​‌ the IFToMM (International Federation​​ for the Promotion of​​​‌ Mechanism and Machine Science)​ technical Committees on History​‌ and on Computational Kinematics.​​ He is a member​​​‌ of the IFToMM Executive​ Council Publication Advisory Board​‌ and an IEEE Fellow.​​
• Jean-Pierre Merlet is a​​​‌ member of the advisory​ "Conseil des Sages" of​‌ GDR Robotique and participates​​ to the GDR working​​​‌ group on “Frugality and​ sobriety”, whose purpose is​‌ to reduce the ecological​​ impact of robotics.

9.1.5​​​‌ Scientific expertise

• Jean-Pierre Merlet​ is a nominator for​‌ the Japan’s Prize and​​ participated to the WAICF​​​‌ conference in Cannes where​ he has presented his​‌ work on neural networks​​ (section 7.1.1).
• Yves​​​‌ Papegay is a member​ of the OpenMath Society,​‌ building an extensible standard​​ for representing the semantics​​​‌ of mathematical objects.
• Eric​ Wajnberg is an appointed​‌ member of the Academic​​ Committee of the Hebrew​​​‌ University of Jerusalem, an​ appointed member of the​‌ International Advisory Board of​​ the “International Center for​​ Excellence in Biological Control”.​​​‌

9.1.6 Research administration

• Yves‌ Papegay is the head‌​‌ of local CUMI (Committee​​ of users of the​​​‌ numerical resources and tools).‌
• Odile Pourtallier is a‌​‌ member of local "Comité​​ de Centre" and of​​​‌ the CUB (Committee of‌ users of the offices)‌​‌
• Within AGOS, the organization​​ in charge of social​​​‌ welfare and employee benefits,‌ Odile Pourtallier serves as‌​‌ the local secretary of​​ the CGL, which is​​​‌ the Local Management Committee‌ of AGOS. At the‌​‌ national level, she is​​ a member of the​​​‌ executive board and is‌ responsible for the AGOS‌​‌ Archives mission. This mission​​ consists in defining the​​​‌ future of the AGOS‌ archives since its creation,‌​‌ which are currently stored​​ in Rocquencourt. Part of​​​‌ these archives will be‌ used to document and‌​‌ write the history of​​ AGOS from its origins.​​​‌ Odile Pourtallier is also‌ responsible for the ‘AGOS‌​‌ Benefits’ working group. This​​ group is in charge​​​‌ of redefining and harmonizing‌ the subsidy and benefit‌​‌ policies implemented by all​​ local CGLs, in order​​​‌ to ensure their consistency‌ and their full compliance‌​‌ with URSSAF regulations.

9.2.1 Supervision‌​‌

• Jean-Pierre Merlet and Yves​​ Papegay are supervising together​​​‌ the PhD of Clara‌ Thomas , who just‌​‌ completed her second year​​ by working on the​​​‌ use of the modularity‌ of CDPR for green‌​‌ and mobility assistance robots.​​

9.2.2 Juries

• Jean-Pierre Merlet​​​‌ has been a member‌ of a PhD defense‌​‌ jury at Paris Sorbonne​​ University (Lê A., 2025/10/15)​​​‌
• Yves Papegay has been‌ a member of a‌​‌ PhD defense jury at​​ Politech Angers (Remin, H.,​​​‌ 2025/01/10)

9.2.3 Educational and‌ pedagogical outreach

• Jean-Pierre Merlet‌​‌ has been one of​​ the lecturer at the​​​‌ 5th Summer School on‌ Singularities of Mechanisms and‌​‌ Robotic Manipulators.

9.3.1 Productions (articles, videos,​​​‌ podcasts, serious games, ...)‌

• Jean-Pierre Merlet prepared an‌​‌ introductory tutorial to robotics​​ using the ePoc mobile​​​‌ learning format, which is‌ currently under test with‌​‌ end-users.

9.3.2 Participation in​​ Live events

• In the​​​‌ scope of Inform@thiques.fr, Yves‌ Papegay organized several School‌​‌ on Experimental Mathematics in​​ July in Oxford, for​​​‌ high-school students, during high-school‌ holidays.
• Yves Papegay animated‌​‌ two master classes (in​​ French and in Romanian)​​​‌ for high school students,‌ one in Spring in‌​‌ Pertuis (France) on Neural​​ Networks, and one in​​​‌ Autumn on generative AI‌ in Cluj-Napoca (Romania).

International journals

• 14​ articleD.Diandra Achre​‌, E.Eric Wajnberg​​ and F. L.Fernando​​​‌ Luís Cônsoli. Gut​ bacteria of Spodoptera frugiperda​‌ establish endophytic association and​​ affect the interactions of​​​‌ their host herbivore with​ maize plants.Journal​‌ of Pest Science98​​2January 2025,​​​‌ 1003-1018HALDOIback​ to text
• 15 article​‌ T.Tamar Keasar,​​ M.Moshe Coll,​​​‌ Y.Yael Lubin,​ M.Michal Segoli,​‌ S.Saskya van Nouhuys​​, E.Eric Wajnberg​​​‌ and A.Alon Tal​. Judiciary independence: Why​‌ should conservation ecologists care?​​ Conservation Science and Practice​​​‌ 7 4 March 2025​ HAL DOI back to​‌ text
• 16 articleT.​​Tamar Keasar and E.​​​‌Eric Wajnberg. What​ maintains variation in flower​‌ accessibility to pollinators in​​ plant communities? A simulation​​​‌ study.BMC Ecology​ and Evolution251​‌May 2025, 45​​HALDOIback to​​ text

National journals

• 17​​​‌ articleA.-V.Anne-Valérie Gasc‌ and Y.Yves Papegay‌​‌. Le voisinage comme​​ Jeu de la vie​​​‌.Design, Arts, Médias‌June 2025HALback‌​‌ to text

International peer-reviewed​​ conferences

• 18 inproceedingsJ.-P.​​​‌Jean-Pierre Merlet. Estimating‌ the Young modulus of‌​‌ cables material in cable-driven​​ parallel robots.Mechanisms​​​‌ and Machine Science (Springer)‌CableCon 2025 - 7th‌​‌ International Conference on Cable-Driven​​ Parallel Robots182Proceedings​​​‌ of the 7th International‌ Conference on Cable-Driven Parallel‌​‌ RobotsHong-Kong, ChinaSpringer​​July 2025HALDOI​​​‌back to text

Scientific‌ books

• 19 bookE.‌​‌Eric Wajnberg and M.​​Michal Segoli. Life​​​‌ History Evolution Traits, Interactions,‌ and Applications.Wiley‌​‌2025HALback to​​ text