Despite significant advances in bipedal locomotion, enabling humanoid robots to perform general whole-body tasks through meaningful interaction with their environments remains a challenging open problem. While deep reinforcement learning (RL) has recently demonstrated impressive results in dynamic walking — even on complex and unpredictable terrain — real-world utility demands that humanoids go beyond locomotion to execute task-oriented behaviors.In this work, we propose a framework for teaching humanoid robots to imitate humans doing useful tasks by training policies for tracking human motion references. Our approach leverages high-quality in-house motion capture (MoCap) data, from which we perform kinematic retargeting to project human trajectories onto a humanoid platform. Crucially, we adopt a hybrid learning paradigm: the policy is trained to track upper-body and root motions from the MoCap data, and receives additional supervision from a pre-trained omnidirectional walking expert. This expert guidance, implemented via a Behavior Cloning (BC) objective, ensures that leg motion respects dynamics and kinematic constraints of the humanoid. We train policies entirely in simulation and successfully transfer them to a real humanoid robot. We validate our method on a box loco-manipulation task, demonstrating effective sim-to-real transfer and marking a step toward more capable, task-driven humanoid behavior.

Deep Reinforcement Learning (RL) is now commonly used for controlling legged robots. Several recent studies have demonstrated impressive results in solving increasingly complex robotic tasks such as navigation in unstructured environments or loco-manipulation. However, this complexity often comes with intricate learning setups requiring tedious reward shaping and features to help convergence. In this work, we tackle these issues and achieve loco-manipulation with a humanoid robot using a RL algorithm that enforces constraints through stochastic terminations during policy learning. We keep the number of rewards low by reformulating them as constraints when they can be intuitively expressed that way. Moreover, we study the relevance of various learning features encountered in the literature and show that providing observations without noise or privileged information to the critic are two straightforward ways to boost locomotion performances on rough terrains. We also demonstrate that the proposed minimalist architecture is not limited to pure locomotion but extends to a loco-manipulation task involving upper limbs. Videos are available at humanoid-cat.github.io.

We propose a novel walking control scheme based on the dynamics of the Linear Inverted Pendulum (LIP) model. The pattern generation incorporates a model of contact forces, enabling closed-loop control of the humanoid robot’s state, including the Center of Mass (CoM) position, velocity, and Zero Moment Point (ZMP). No additional control policies are required to maintain static and dynamic balance. Our approach also includes dynamic re-planning of step locations and timings, thus preserving the LIPs boundedness condition. We validated this controller on five different humanoid robots, testing its ro- bustness through various disturbances, including sudden pushes during walking and static phases. Additionally, our controller demonstrated effective locomotion over uneven and compliant terrain. Both simulation and experimental results confirm the effectiveness and robustness of this controller.

This paper presents a modular manipulation pipeline for CALL-M, a mobile robot developed at CNRS-AIST JRL for semi-autonomous pick-and-place operations in convenience store environments. The system leverages a ROS 2-based architecture integrating 3D perception, grasp detection, and motion planning using MoveIt 2. The pipeline comprises modular stages—point cloud acquisition, object selection, grasp estimation, and trajectory generation—coordinated by a centralized task manager.Validation in both simulation and real-world scenarios demonstrated successful grasps. While simulation confirmed reliability under ideal conditions, real-world trials revealed challenges due to sensor noise, workspace constraints, and misalignments in grasp pose generation. Despite this, the system’s modularity and adaptability make it a scalable solution for manipulation in semi-structured environments.

Camera models play a crucial role in robot vision applications. Yet their diversity poses a challenge when working with data captured with cameras calibrated using different models. In this paper, we address this issue by introducing a mathematical framework that enables conversion between various camera projection models. This approach allows algorithms designed for a specific model to process data from cameras calibrated with other models, eliminating the need for recalibration and enabling the reuse of pre-existing datasets that do not provide access to calibration images. We present the general conversion method for state-of-the-art camera models that we derive for three new camera model conversions, covering various camera types, including fisheye and catadioptric systems. Quantitative evaluation is conducted with respect to well-known calibration methods. We compare our method on image undistortion, as well as in practical applications such as SLAM, visual servoing, and visual odometry. The results demonstrate that our conversion approach achieves performances comparable to calibration without the need for explicit calibration. This work contributes to a more flexible and adaptive use of cameras in robot applications. The proposed camera model conversion framework is implemented in the open-source libPeR library.

This paper investigates an estimation-based force control approach for humanoid robots’ body parts that lack direct force sensors. We employ a "Kinetics Observer", originally developed for legged robots, which can estimate various robot states, including external forces. In addition, we utilize a "Multi-Contact Controller" for whole-body balance control under multiple contact conditions. The estimated external force is used as feedback to control the humanoid’s whole-body balance, partially without relying on a force sensor. For verification, we used hands that are typically equipped with six-axis force sensors; while we did not use the sensor feedback in the proposed control scheme. The experiment results show that estimation-based force control can replace sensor-based control at least for static balancing.

In this paper, we propose the multi-contact locomotion planning framework for humanoid robots that leverages the target contact selection considering its long-term preferability by reinforcement learning (RL) in optimization-based motion generation with feasibility constraints. It is difficult to predict how the next target contact will affect the motion of the robot over the future in multi-contact locomotion, where humanoid robots are required to perform complex motions with kinematic constraints and static equilibrium. To solve this problem, we evaluate the preferability of the motion planned by the optimization-based motion planner to reach the target contact, which we define as contact preferability, as the reward for RL. This idea enabled us to train the policy to provide a contact with the large future preferability without explicitly designing its future promise by ourselves. We also propose the design of action space for RL based on the robot's reachability. We construct sets of feasible joint angles for each limb of the robot as successors and use them as the action space instead of directly managing contact poses. By defining the deterministic mapping from the successor to the target contact, the proposed framework can manage acyclic multi-contact motion where the number of contacts can be changed. We evaluate the proposed framework in three scenarios and prove that it can plan a preferable contact sequence for multi-contact locomotion with a high success rate and short computational time.

To reduce the computational cost of humanoid motion generation, we introduce a new approach to representing robot kinematic reachability: the differentiable reachability map. This map is a scalar-valued function defined in the task space that takes positive values only in regions reachable by the robot's end-effector. A key feature of this representation is that it is continuous and differentiable with respect to task-space coordinates, enabling its direct use as constraints in continuous optimization for humanoid motion planning. We describe a method to learn such differentiable reachability maps from a set of end-effector poses generated using a robot's kinematic model, using either a neural network or a support vector machine as the learning model. By incorporating the learned reachability map as a constraint, we formulate humanoid motion generation as a continuous optimization problem. We demonstrate that the proposed approach efficiently solves various motion planning problems, including footstep planning, multi-contact motion planning, and loco-manipulation planning for humanoid robots.

3D point cloud registration is paramount for mapping complex environments. However, aligning overlapping clouds remains computationally demanding due to the increasing size and density of data from modern Terrestrial Laser Scanners. In this paper, we propose a lightweight, memory-efficient registration approach that operates directly on a coarse level of detail extracted from octree structures. Rather than relying on keypoints, we construct full coarse representations and associate each point with a descriptor to form global feature maps. We evaluate several point distribution strategies for building these representations and show that our approach significantly reduces processing time and memory usage while maintaining or improving registration accuracy.

This article describes RHP Friends, a social humanoid robot developed to enable assistive robotic deployments in human-coexisting environments. As a use case application, we present its potential use in nursing by extending its capabilities to operate devices and tools according to the task and by enabling remote assistance operations. To meet a wide variety of tasks and situations in environments designed by and for humans, we develop a system that seamlessly integrates the slim and lightweight robot and several technologies: locomanipulation, multicontact motion, teleoperation, and object detection and tracking. We demonstrate the system's usage in a nursing application. The robot efficiently performs the daily task of patient transfer and a nonroutine task, represented by a request to operate a circuit breaker. This demonstration, held at the 2023 International Robot Exhibition (IREX), was conducted three times a day over three days.

In this paper, we propose a novel approach for efficient 2D navigation using a multimodal sensor fusion technique. Our method focuses on merging data from multiple sensors, such as LiDARs, cameras, and ultrasonic sensors, into a unified Laser-Scan, which serves as a foundation for faster and more lightweight navigation. By fusing sensor data at the Laser-Scan level, our approach enables the use of basic 2D Simultaneous Localization And Mapping (SLAM) algorithms for mapping tasks, or any others Laser-Scan based features, while still benefiting from the rich information provided by multimodal 3D inputs. This results in a more computationally efficient solution compared to traditional 3D methods that rely on depth points or full multimodal SLAM systems. Our experimental results demonstrate that the proposed approach achieves comparable accuracy in mapping and localization while significantly reducing computational complexity and processing time. This research offers a promising alternative for real-time 2D navigation in resource-constrained autonomous systems, such as drones or any small unmanned vehicles.

This work explores the integration of RGB-D and hyperspectral imaging technologies to enhance the accuracy of 3D reconstructions. We propose a methodology that combines depth maps from an RGB-D camera with spectral data from a hyperspectral camera to develop a detailed 3D hyperspectral point cloud model. Our approach combines feature points detected in spectral bands and captured depth maps to estimate the poses of the multi-camera system moved around objects to reconstruct their overall shape accurately in a dense manner. 3D reconstruction with the hyper-spectral data experiments are conducted on commonly used objects in literature and compared with the publicly available ground-truth 3D point clouds of these objects. Results of utilizing full spectral information consistently demonstrate superior reconstruction accuracy both quantitatively and qualitatively compared to reduced spectral representations (i.e. three band selected false-color images or band averaged images).

In this paper, we propose the "Kinetics Observer", a novel estimator addressing the challenge of state estimation for legged robots using proprioceptive sensors (encoders, IMU and force/torque sensors). Based on a Multiplicative Extended Kalman Filter, the Kinetics Observer allows the real-time simultaneous estimation of contact and perturbation forces, and of the robot's kinematics, which are accurate enough to perform proprioceptive odometry. Thanks to a visco-elastic model of the contacts linking their kinematics to the ones of the centroid of the robot, the Kinetics Observer ensures a tight coupling between the whole-body kinematics and dynamics of the robot. This coupling entails a redundancy of the measurements that enhances the robustness and the accuracy of the estimation. This estimator was tested on two humanoid robots performing long distance walking on even terrain and non-coplanar multi-contact locomotion.

To efficiently achieve complex humanoid loco-manipulation tasks in industrial contexts, we propose a combined vision-based tracker-localization interplay integrated as part of a task-space whole-body optimization control. To achieve good perception complementarity between manipulation and localization, a new fast dense 3D model-based tracking using wide-angle depth image is developed and used in conjunction with a simultaneous localization and mapping software. Our approach allows humanoid robots, targeted for industrial manufacturing, to manipulate and assemble large-scale objects while walking. It is assessed with experiments consisting in rolling and assembling in an unwinder a heavy and wide bobbin using bimanual grasping and bipedal locomotion at a time. This experimental use-case is found in some large-scale manufacturing where bobbins are enrolled with various materials (cables, papers, rubbers, etc.). The same experiments are made using two different humanoid robots of the same family. Note to Practitioners-This paper aims at deploying humanoid robots in large-scale manufacturing industries. We consider nonadded value tasks related to transporting large tools or objects such as large bobbins by means of locomanipulation skills, similarly to human workers. We developed a task-space control framework that has been successfully applied in the aircraft industry. In the frame of a current collaboration with other major industrial sectors, we enhanced our control framework to interplay between SLAM and visual tracking to realize robust loco-manipulation tasks. Our approach can be applied and ported to any humanoid robot or bi-manual wheeled mobile robots with minor programming effort as the software is made open. Preliminary experiments with two different humanoids and use-cases suggest that our approach is feasible. In future research, we will address the problem of performance to reach at least human-speed in the execution of locomanipulation tasks in large-scale industry and automation contexts.

To enable humanoid robots to work robustly in confined environments, multi-contact motion that makes contacts not only at extremities, such as hands and feet, but also at intermediate areas of the limbs, such as knees and elbows, is essential. We develop a method to realize such whole-body multi-contact motion involving contacts at intermediate areas by a humanoid robot. Deformable sheet -shaped distributed tactile sensors are mounted on the surface of the robot's limbs to measure the contact force without significantly changing the robot body shape. The multi-contact motion controller developed earlier, which is dedicated to contact at extremities, is extended to handle contact at intermediate areas, and the robot motion is stabilized by feedback control using not only force/torque sensors but also distributed tactile sensors. Through verification on dynamics simulations, we show that the developed tactile feedback improves the stability of whole-body multi-contact motion against disturbances and environmental errors. Furthermore, the life-sized humanoid RHP Kaleido demonstrates whole-body multi-contact motions, such as stepping forward while supporting the body with forearm contact and balancing in a sitting posture with thigh contacts.

This paper presents a software system designed for indoor quadrupedal locomotion using vision-based techniques. It incorporates established technologies such as Simultaneous Localization and Mapping (SLAM), global planning, and composite Center of Gravity (COG) trajectory planning methods to achieve both autonomous and semi-autonomous quadrupedal movement. The software solution offers a comprehensive approach to locomotion challenges, encompassing vision-based foothold planning and obstacle avoidance. With a modular architecture that supports diverse controller implementations, the software demonstrates its versatility. Overall, this research contributes a comprehensive software framework for indoor quadrupedal locomotion, exhibiting promising results in simulation and offering potential for real-world deployment.

For the deployment of legged robots in real-world environments, it is essential to develop robust locomotion control methods for challenging terrains that may exhibit unexpected deformability and irregularity. In this paper, we explore the application of sim2real deep reinforcement learning (RL) for the design of locomotion controllers for large-sized humanoid robots on compliant and uneven terrains. Our key contribution is to show that a simple training curriculum for exposing the RL agent to randomized terrains in simulation can achieve robust walking on the real humanoid robot using only proprioceptive feedback. We train an end-to-end omnidirectional locomotion policy using the proposed approach and show extensive real robot demonstration on the HRP-5P humanoid over several difficult terrains inside and outside the lab environment.

Almost all (not to say all) existing humanoids have not been designed with the ability to change their end-effectors (head, feet and hands) on-the-fly. Inspired by the tool-changing mechanisms in robotics automation and manufacturing, we propose an end-effectors changer mechanism suitable to humanoids. This letter explains why enabling humanoids with such a technology is important and why existing tool-changer mechanisms are not adapted. The proposed changer mechanism is not actuated. Yet, it does not require human intervention to assist the change of end-effectors. We assess our mechanism through a comparative study with existing tool-changers and demonstrations with the HRP-4 humanoid. We claim that our idea could be a new turn in the design of future humanoids and open perspective to modular sizing of robots. The proposed mechanism can possibly apply to animaloids to some extent.