Whole-body planning and control consist of techniques that exploit the entire body structure of a robot, its redundancy and its environment to execute a desired movement. Achieving complete whole-body control of a robotic system with real-time performances increases the capability of the robot to complete complex loco-manipulation tasks, namely in terms of agility and dexterity.
Multi-contact and agile loco-manipulation
The first objective of HUCEBOT is to design whole-body control schemes and algorithms that enable humanoid robots, exoskeletons or digital human models to control their movements while interacting with the environment (including humans) engaging multiple contacts, manipulating heavy payloads, and executing agile and dynamic tasks. Those challenges, although consequent for both humanoid robots and exoskeletons, are necessary to create robots that can physically help humans. Robots need to produce suitable forces and follow the humans in their movements, which are notably faster and more agile than what robots are capable of doing today.
Our ambitious goal is to design and deploy whole-body control schemes on our robots to achieve the execution of agile loco-manipulation and teleoperated whole-body manipulation which include actions such climbing stairs, carrying payloads, pushing carts and manipulating objects in the environment at the natural human speed. In order for robots to assists humans, they need to keep the pace that humans have in their activities, and not slow them down to the point of frustration and human rejection. To this end, we need to enable the robots to move faster, execute more agile and dynamic tasks, and deal with high payloads.
Leveraging machine learning for modeling, control and agile motions
HUCEBOT investigates complementary solutions that leverage data-driven learning, both for generating candidate trajectories (Jelavic et al. 2023) and to explore new controllers that go beyond the limits imposed by model-based controllers. Our goal is to increase the number of degrees of freedom while applying more sophisticated approaches by focusing our efforts on contact-rich motion and minimizing the amount of “reward hacking” required to get effective controllers. We use machine learning to improve our current contact models, with data-driven contact models to better represent the complex physics of the interaction between robot parts, the environment, and the human body, which is soft, deformable, and user-specific. HUCEBOT takes interest in computational models that can be incorporated into real-time control and potentially embedded on wearable devices. Our team also focuses on rapid reconfigurations of the humanoid posture through which increased speed of humanoid robot’s contact improves task manipulability while reducing their energy consumption and maintaining their balance.
Human-aware collaborative controllers
HUCEBOT aims to design collaborative controllers that are “human-aware”: at a low level, this means that the whole-body controller considers the human dynamics, possibly anticipating the human intents. Past work highlighted some limits which we are currently looking into by tailoring our approach to reason in probabilistic terms about the possible ways humans will act and react to the robot’s behavior. To anticipate the human and adapt itself accordingly, the robot should use a prediction of the human’s intention and plan consequently.
