Skin and whole-body motion of large objects with torque and position controlled humanoid robots – HoRoPo

High-Level Planning

High-level planning in the context of this project could have involved sequencing contacts to manipulate an object—for example, planning how to flip an object (Lamiraux et al, 2021). However, implementing such a mechanism requires a system capable of generating continuous motion for a given sequence of contacts. Within the Gepetto group, we have developed an approach based on whole-body predictive control over many years. The associated optimization problem is solved using Differential Dynamic Programming (DDP) via the Crocoddyl software [CMB20]. To explore the use of Crocoddyl on the TALOS robot, we initially limited contact modeling to skin patches. We also had to account for the fact that, since the system is torque-controlled, actuator dynamics can become unstable with respect to gains derived from sensitivity analysis.

Low-Level Control of Interaction Forces at Contact Points

Low-level force control was initially intended to rely on Riccati gains derived from the whole-body predictive controller. Theory suggests that between two optimization problems corresponding to successive iterations of Model Predictive Control (MPC), certain gains from the DDP algorithm allow for calculating the Taylor expansion between the two solutions [AVF83]. Verifying this hypothesis proved particularly important for subsequent work.

Another critical aspect was the partial calibration of the sensors provided by InTouch Robotics: they were calibrated to provide consistent output but not to estimate force. Part of our work therefore involved developing a method to calibrate these forces. Finally, to model contacts, it is essential to simulate interactions with the environment.

Estimation of Unknown Object Parameters Using Tactile and Joint Sensors

We also worked on estimating the parameters of an object when manipulated by a robot equipped with artificial skin. This involves first estimating its weight and center of mass, followed by its inertial parameters. We primarily leveraged the unique capabilities of our TALOS robot, equipped with torque sensors, enabling it to measure forces applied to the motors and the data provided by the skin.

Another critical aspect was the partial calibration of the sensors provided by InTouch Robotics: they were calibrated to provide consistent output but not to estimate force. Part of our work therefore involved developing a method to calibrate these forces. Finally, to model contacts, it is essential to simulate interactions with the environment.

Estimation of Unknown Object Parameters Using Tactile and Joint Sensors

We also worked on estimating the parameters of an object when manipulated by a robot equipped with artificial skin. This involves first estimating its weight and center of mass, followed by its inertial parameters. We primarily leveraged the unique capabilities of our TALOS robot, equipped with torque sensors, enabling it to measure forces applied to the motors and the data provided by the skin.

One of the major outcomes of the project was the presentation by Rajesh Subburaman (a postdoctoral researcher on the project) of the paper "Delay Robust Model Predictive Control for Whole-Body Torque Control of Humanoids" at the international Humanoids 2024 conference. This presentation was selected as a finalist for the best presentation award. It highlighted the importance of managing several critical factors that must be taken into account for the effective implementation of whole-body controllers on humanoid robots. New results are expected to improve the calibration of force sensors derived from this technology.

Scientific Output and Patents Since the Project’s Inception:

No patents have been filed as part of this project. To date, the scientific output consists of two conference papers published in the proceedings of the leading humanoid robotics conference. The first paper clarified key aspects necessary for implementing a whole-body controller on the TALOS humanoid robot. The second paper addressed two main points: estimating object parameters (such as the position of the center of mass and inertial parameter identification) and calibrating the skin sensors. A journal article and an additional conference paper are expected to be submitted after the project’s conclusion.

The project faced significant delays due to technical issues with the aging TALOS robot, which is also a prototype. Nevertheless, we have substantial material for high-quality journal and conference papers on state estimation and simulation.

The project enabled us to achieve several objectives, particularly related to the integration of artificial skin on the TALOS humanoid robot and the implementation of whole-body torque predictive control. This approach proved particularly complex for several reasons. First, the complexity of the problem itself can lead to variable solving times, making it crucial to manage the resolution delays effectively. Additionally, because it involves torque control, the dynamics of the actuators must be carefully considered. This challenge was partially addressed in the first conference paper, which received an award.

From a purely technical standpoint, the TALOS robot available at LAAS is the first of its kind and suffers from design flaws, particularly with its connectors. This makes experiments on the robot especially difficult. Fortunately, the project benefited from close collaboration between LAAS and PAL Robotics through the joint laboratory Dynamograde, which provided continuous support from a PAL Robotics engineer. While this support was not funded by the project, none of the presented work would have been possible without it.

From a scientific and technical perspective, the project will also benefit from the work carried out under the AGIMUS project. All the work done by Rajesh Subburaman was conducted using ROS-1, which is no longer supported. However, a significant portion of the code has already been ported, and transitioning to ROS-2 should be relatively straightforward. This was the purpose of the evaluation work conducted on the Gazebo Harmonic simulator, particularly with the DART simulator. It aligns with the detailed analysis performed by the INRIA WILLOW team on contact modeling using Alternating Direction Method of Multipliers (ADMM) approaches. DART offers similar stability using Gauss-Seidel projectors.

In parallel, the LAAS GEPETTO team explored reinforcement learning-based control methods through other funding sources. These methods are scientifically closely related, except for the memory aspect of neural networks. This project serves as an excellent foundation for future projects leveraging the sense of touch to build Vision-Language-Action (VLA) models, incorporating contact information and action examples provided by whole-body predictive control.

The goal of HoRoPo is to develop a new scientific paradigm for motion generation considering humanoid robots equipped with a skin. The aim is to manipulate large objects using torque-controlled and non-torque-controlled humanoid robots and comparing the limitations of the two approaches and to fuse them to a uniformed controller.
The scientific direction will be to develop model predictive control strategies to take advantage of the skin, and the torque- and non-torque control. A scientific question to solve is the complexity arising from the number of contacts created by the object on the robot. HoRoPo will explore and exploit the potential of the skin for full-body manipulation control. It will take its foundation on the work by the German team (TUM-ICS) on the skin and the French team (LAAS-Gepetto) on whole-body predictive control. The former has realized some impressive breakthroughs in making the transmission of contact on skin patches accessible and manageable. The latter has proposed the new algorithm Crocoddyl for model predictive control for computing whole body motion for a humanoid and make MPC feasible for a torque-controlled robot with state feedback. TUM-ICS has already equipped REEM-C, a full-size humanoid robot (H1), with skin and generated whole body motion using tactile feedback without torque-control. Five years ago, LAAS-Gepetto has proposed improvements with respect to the REEM-C humanoid, in order to have a humanoid which is quite powerful and can be controlled at the torque level for the actuators. The ambition of HoRoPo is to equip a torque-controllable robot (TALOS) with skin to manipulate heavy objects and investigate a comparable strategy for non-torque-controllable humanoids ( H-1).
LAAS-Gepetto will extend the formulation of Croccodyl to handle multiple contacts, and modify on-line the foreseen behaviour according to the robot state. One difficulty with whole body MPC is the richness of the possible trajectory, since the algorithm can be trapped in unfeasible local minima. To cope with this, a successful approach is to use Machine Learning to learn the most likely solution for the problem through a statistical exploration of the problem.
The prototype robotic skin developed at TUM-ICS provides multi-modal sensing capability, pressure, temperature, acceleration. It has already been used to cover their humanoid H-1 for research on balancing and walking control, including a software framework. In HoRoPo, we will join force with LAAS to work on extending this capability to whole-body manipulation. TUM-ICS will focus on contacts detection and multi-contacts-based control of a humanoid robot with position or torque control capabilities.
The new control schemes developed in HoRoPo will open up a new line of humanoid robotic research. Such advanced technologies will have an impact which will go far beyond humanoid robotics to date and will increase the use of these robots across multiple application domains.