The scientific heritage of Bleu Robotics

Introduction

Over the past twenty years, the founding team of Bleu Robotics and their team from Inria/CNRS/Univ. Lorraine (HuCeBot team) have developed algorithms for humanoids that learn and adapt. A guiding principle throughout this work has been that learning should be applied where it makes a concrete difference for robotics, so that both model-based and data-driven approaches complete each other instead of competing.

This work is organized around three questions:

1. Whole-body control: how to synchronize 20-30 motors to both achieve the task and respect the constraints, like keeping the robot balanced?
2. Data-efficient learning: how to learn with very few trials or very few demonstrations, like 10-30 trials / a few minutes?
3. Humanoid learning: how to leverage these algorithms in humanoid robots?

Part I - Whole-Body Control for Humanoid Robots

The first challenge in humanoid robotics is getting the robot to move at all. A humanoid has 20 to 30 actuated degrees of freedom — hips, knees, ankles, spine, shoulders, elbows, wrists — and every motion requires synchronizing all of them simultaneously. Moving a hand to a desired position while keeping the robot balanced is not a simple inverse-kinematics problem; it demands reconciling dozens of coupled constraints in real time.

After many years of research, the humanoids community converged on whole-body control based on quadratic programming (QP). At each control cycle (200–1000 Hz) the controller solves an optimization problem:

where $\tau^{\ast }$ and $\ddot{q}^{\ast }$ are the optimal joint torques and accelerations. $w_k$ is the weight of the task $k$ described by $A_k(q,\dot{q})$ and $b_k(q,\dot{q})$ . $M(q)$ is the joint-space mass matrix, $h(q,\dot{q})$ contains all the non-linear terms (Coriolis, centrifugal and gravity), $J(q)$ is the contact Jacobian and $f$ the contact forces. Equality constraints are described by $A_{eq}(q,\dot{q})$ and $b_{eq}(q,\dot{q})$ . Inequality constraints are described by $A_{ineq}(q,\dot{q})$ and $b_{ineq}(q,\dot{q})$ . Tasks are weighted on the same level and their priority directly depends on $w_k$ .

We get the optimal joint accelerations and integrate them twice to control the robot in position, although we could use the same values and control the robot directly in torque.

This QP architecture is a major achievement of model-based robotics, providing real-time, physically consistent control with hard safety guarantees. Our early humanoid work was concerned with making this framework more capable, more robust, and more practical to deploy.

Motion retargeting from human to humanoid

Our first contribution addressed the problem of whole-body motion retargeting: transferring a human operator's movements to a humanoid in real time while respecting the robot's physical constraints [1]. This is both useful for teleoperating remote humanoids and to record whole-body demonstrations.

The core difficulty is that humans and humanoids have different kinematics — limb proportions, joint ranges, and balance dynamics all differ — so direct joint-angle mapping produces infeasible or unstable motions. We developed a retargeting pipeline that maps motion-capture data to task-space references on the fly, then feeds them into the QP whole-body controller to resolve conflicts between the retargeted motion and balance.

Whole-body teleoperation

Building on this retargeting foundation, we developed a complete whole-body teleoperation system for the Talos humanoid [2]. The system maps a human operator's full-body motion to task-space references tracked in real time by the QP controller, with balance maintained throughout. Because motion mapping and QP execution are integrated into a single pipeline running at control frequency, an operator can intuitively command complex whole-body behaviors — simultaneous loco-manipulation, transitions between support configurations — without any knowledge of the robot's internal controller. These preliminary results demonstrated the feasibility of high-bandwidth, whole-body humanoid teleoperation and identified the key bottlenecks for further work.

Multi-contact force control for position-controlled robots

Humans often use additional contact points to enhance their stability, for instance, by using a handrail or a wall when walking, or to extend their reach, for instance, when grasping a distant object. While humanoid robots would benefit from a similar strategy, current robots minimize the number of contacts and use them only for feet and required interactions with the environment, such as pushing a button. The primary challenge in controlling multi-contact lies in the redundancy of force distribution resulting from closed kinematic chains. For a given posture with several contacts, there are infinite ways to distribute force among them. For instance, a humanoid with both hands on a table can apply more or less force to the hands without any visible change in joint position.

We developed a whole-body multi-contact controller specifically designed for position-controlled humanoids (instead of torque-controlled) of robots [3]. By augmenting the QP with compliance models and estimating contact forces from joint encoders, the controller can track desired force profiles at multiple simultaneous contacts — without any additional force/torque sensors — while maintaining balance and respecting joint limits. This closes the loop on the basic whole-body control pipeline: from retargeting human motion, to teleoperating the full body, to explicitly regulating the forces exchanged with the environment.

Part II - Data-Efficient Learning

If learning is to be useful on physical robots, it must work with very little data. A robot cannot afford thousands of trials to learn a behavior — hardware is fragile, time is limited, and the environment may be changing. This constraint has shaped our approach from the outset: we design every algorithm to extract maximum information from minimal interaction.

One of our most impactful contributions is the Intelligent Trial and Error (IT&E) algorithm, which we published in Nature [4] (and was selected for the cover). Our objective was to allow robots to adapt to damage without a diagnosis, but by trial-and-error learning. The method operates in two phases: in the offline phase, we use a MAP-Elites quality-diversity algorithm to generate a behavioral repertoire — a discrete map of thousands of high-performing gaits indexed by behavioral descriptors (e.g., the fraction of time each leg is in contact with the ground). When the physical robot is damaged, the online phase uses Bayesian optimization over this repertoire to find a compensatory gait, typically in under two minutes and fewer than twenty trials. The repertoire acts as a structured prior that dramatically reduces the search space — an example of our general strategy of investing computation offline (in simulation) so that online learning on the real robot can be fast. We demonstrated recovery from six different damage conditions on a six-legged robot, including missing and shortened legs.

As a side note, this contribution introduced the MAP-Elites algorithm, which is now used in numerous fields beyond robotics, like for tuning LLMs (Llama 2.1 (Meta)) or in Alpha Evolve to find solutions to scientific problems (Google).

We then pursued a major line of follow-up work on how to construct, select, and use these priors more effectively. We parameterized the GP prior mean with a low-fidelity simulator to exploit approximate dynamics while correcting for sim-to-real discrepancies [6]. We proposed a method to select among candidate priors automatically during optimization [5], and we extended this to online selection among multiple behavioral repertoires [9]. In parallel, we removed the common assumption that the robot can be reset between trials, enabling continuous damage recovery without interruption [7], and we introduced a novel black-box data-efficient policy search [8]. In each case, our motivation was the same: make learning practical under the constraints of real hardware, where every trial counts.

We also brought these data-efficient principles to manipulation [10], where we learned object-centric grasp preferences from a small number of demonstrations and integrated these into a grasp planner. Rather than attempting to learn grasping from scratch, we learn only what is hard to pecify manually — the preferences — and delegate the rest to a conventional planner.

Part III - Learning for Humanoid Robots

Humanoid robots present specific challenges that go beyond those of legged or manipulator platforms. The QP whole-body controller described above provides real-time, physically consistent control with hard safety guarantees. We do not seek to replace it with learning. Instead, we use learning to solve the problems that the QP framework leaves open — and there are many. The QP controller answers "how to execute" but not "what to execute," "with which parameters," or "under which constraints." We focus on these gaps.

Automatic controller tuning

The first and most immediate gap is parameter tuning. A QP whole-body controller exposes task weights, priority orderings, feedback gains, and constraint margins that interact in non-obvious ways. We used CMA-ES (single-objective) and NSGA-II (Pareto-based multi-objective) to optimize these parameters on the iCub humanoid, evaluating performance across randomized scenarios — replacing hours of manual tuning with an automated search that produces more robust gains [11]. We then generalized this by treating the controller structure itself — which tasks to include, their ordering, their parametrization — as part of the search space, and we automatically selected and tuned complete whole-body controller configurations on the Talos humanoid [12]. In both cases, we applied learning to the specific bottleneck (parameter and structure selection) while the QP framework continued to provide real-time execution and constraint satisfaction.

Learning to deal with damage on a humanoid

Humanoid robots could replace humans in hazardous situations but most of such situations are equally dangerous for them, which means that they have a high chance of being damaged and falling. We hypothesize that humanoid robots would be mostly used in buildings, which makes them likely to be close to a wall. To avoid a fall, they can therefore lean on the closest wall, as a human would do, provided that they find in a few milliseconds where to put the hand(s). We introduced a method, called D-Reflex, that learns a neural network that chooses this contact position given the wall orientation, the wall distance, and the posture of the robot. This contact position is then used by a whole-body controller to reach a stable posture. We show that D-Reflex allows a simulated TALOS robot (1.75m, 100kg, 30 degrees of freedom) to avoid more than 75% of the avoidable falls and can work on the real robot [13].

Prescient teleoperation

While a teleoperated humanoid is, at first sight, a typical example of a non-autonomous robot, it can also benefit from machine learning. For instance, we worked on compensating space-like delays (1 to 2 seconds) for direct teleoperation. To do so, we developed a "prescient teleoperation" system [14] in which a learned predictive model anticipates the operator's future commands, that is perform the commands before having received them. The key insight is to learn to predict the future commands given the previous ones.

Learning manipulation policies with flow matching

With the advances in generative AI, we wondered how to apply modern imitation learning on humanoids. We chose flow matching as the policy representation [15], a continuous flow technique inspired by diffusion that learns to transform a simple noise distribution into the distribution of demonstrated task-space trajectories. Flow matching is particularly well suited to manipulation because demonstrations are inherently multi-modal: there are often several valid ways to grasp an object or sequence a series of contacts, and a standard regression-based policy would average over these modes and produce unusable trajectories. Flow matching preserves the modes and can sample diverse, coherent plans at inference time.

We trained the flow matching policy on teleoperated demonstrations of multi-support manipulation tasks and deployed the resulting policy on a real Talos humanoid — to our knowledge, the first time a flow matching policy was executed on a physical humanoid robot. The learned policy generates task-space references at each control step, which the QP whole-body controller translates into joint commands while enforcing balance, joint limits, and contact constraints. This architecture allowed us to benefit from the expressiveness of flow matching without sacrificing the safety guarantees of the QP layer.

We then extended this approach with Extremum Flow Matching [16], which combines flow matching with offline goal-conditioned reinforcement learning. Rather than imitating the average demonstration, we bias the flow toward the most successful trajectories in the dataset — those that most efficiently reach the desired goal. This allows the humanoid to achieve goals that were only partially or indirectly demonstrated, pushing beyond pure imitation toward goal-directed optimization while retaining the multi-modal expressiveness of flow matching.

Language-guided humanoid control

In our most recent work, we extend this principle to natural language as an input modality. We developed Words2Contact [17], which uses foundation models to identify support contacts from verbal instructions — translating a spoken task description into a contact plan for the whole-body controller. We built on this for safe bimanual teleoperation [18]: we use a language model to process natural-language safety instructions (e.g., "avoid the fragile objects on the left") and translate them into inequality constraints that we inject into the QP controller in real time.

More…

The founding team has together published over 70 journal articles and 95 peer-reviewed conference papers across venues including Nature , IEEE TRO, IJRR, RA-L, ICRA, IROS, and Humanoids, with more than 16,000 citations. A full list of publications is available on the HuCeBot team page, JB Mouret’s and Serena Ivaldi’s academic websites.

1. [1] L. Penco, B. Clément, V. Modugno, E. Hoffman, G. Nava, D. Pucci, N. Tsagarakis, J.-B. Mouret, S. Ivaldi (2018). Robust Real-time Whole-Body Motion Retargeting from Human to Humanoid. IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids) , 425–432.
2. [2] E. Dalin, I. Bergonzani, T. Anne, S. Ivaldi, J.-B. Mouret (2021). Whole-body teleoperation of the Talos humanoid robot: preliminary results. ICRA 2021 — 5th Workshop on Teleoperation of Dynamic Legged Robots in Real Scenarios.
3. [3] Q. Rouxel, S. Ivaldi, J.-B. Mouret (2024). Multi-Contact Whole-Body Force Control for Position-Controlled Robots. IEEE Robotics and Automation Letters .
4. [4] A. Cully, J. Clune, D. Tarapore, J.-B. Mouret (2015). Robots that can adapt like animals. Nature , 521(7553), 503–507. doi:10.1038/nature14422
5. [5] R. Pautrat, K. Chatzilygeroudis, J.-B. Mouret (2018). Bayesian Optimization with Automatic Prior Selection for Data-Efficient Direct Policy Search. IEEE International Conference on Robotics and Automation (ICRA) .
6. [6] K. Chatzilygeroudis, J.-B. Mouret (2018). Using Parameterized Black-Box Priors to Scale Up Model-Based Policy Search for Robotics. IEEE International Conference on Robotics and Automation (ICRA) .
7. [7] K. Chatzilygeroudis, V. Vassiliades, J.-B. Mouret (2017). Reset-free Trial-and-Error Learning for Robot Damage Recovery. Robotics and Autonomous Systems , 1–19. doi:10.1016/j.robot.2017.11.010
8. [8] K. Chatzilygeroudis, R. Rama, R. Kaushik, D. Goepp, V. Vassiliades, J.-B. Mouret (2017). Black-Box Data-efficient Policy Search for Robotics. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) .
9. [9] R. Kaushik, P. Desreumaux, J.-B. Mouret (2020). Adaptive Prior Selection for Repertoire-based Online Learning in Robotics. Frontiers in Robotics and AI .
10. [10] Y. Fleytoux, A. Ma, S. Ivaldi, J.-B. Mouret (2022). Data-efficient learning of object-centric grasp preferences. IEEE International Conference on Robotics and Automation , 6337–6343.
11. [11] L. Penco, E. Mingo Hoffman, V. Modugno, W. Gomes, J.-B. Mouret, S. Ivaldi (2020). Learning Robust Task Priorities and Gains for Control of Redundant Robots. IEEE Robotics and Automation Letters, 5(2), 2626–2633..
12. [12] E. d'Elia, J.-B. Mouret, J. Kober, S. Ivaldi (2022). Automatic Tuning and Selection of Whole-Body Controllers. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) .
13. [13] T. Anne, E. Dalin, I. Bergonzani, S. Ivaldi, J.-B. Mouret (2022). First do not fall: learning to exploit a wall with a damaged humanoid robot. IEEE Robotics and Automation Letters.
14. [14] L. Penco, J.-B. Mouret, S. Ivaldi (2023). Prescient teleoperation of humanoid robots. Proc. of IEEE Humanoids .
15. [15] Q. Rouxel, A. Ferrari, S. Ivaldi, J.-B. Mouret (2024). Flow Matching Imitation Learning for Multi-Support Manipulation. IEEE RAS International Conference on Humanoids .
16. [16] Q. Rouxel, C. Donoso, F. Chen, S. Ivaldi, J.-B. Mouret (2025). Extremum Flow Matching for Offline Goal Conditioned Reinforcement Learning. Proc. of IEEE Humanoids .
17. [17] D. Totsila, Q. Rouxel, J.-B. Mouret, S. Ivaldi (2024). Words2Contact: Identifying Support Contacts from Verbal Instructions Using Foundation Models. IEEE RAS International Conference on Humanoids .
18. [18] D. Totsila, C. Donoso, E. Hoffman, J.-B. Mouret, S. Ivaldi (2025). Safe Bimanual Teleoperation with Language-Guided Collision Avoidance. Proc. of IEEE Conference on Telepresence.