No abstract

No abstract

No abstract

Learning-based methods have proven useful at generating complex motions for robots, including humanoids. Reinforcement learning (RL) has been used to learn locomotion policies, some of which leverage a periodic reward formulation. This work extends the periodic reward formulation of locomotion to skateboarding for the REEM-C robot. Brax/MJX is used to implement the RL problem to achieve fast training. Initial results in simulation are presented with hardware experiments in progress.

Buzzwire tasks are often used as benchmarks and as training environments for fine motor skills and high precision path following. These tasks require moving a wire loop along an arbitrarily shaped wire obstacle in a collision-free manner. While there have been some demonstrations of buzzwire tasks with robotic manipulators using reinforcement learning and admittance control, there does not seem to be any examples with humanoid robots. In this work, we consider the scenario where we control one arm of the REEM-C humanoid robot, with other joints fixed, as groundwork for eventual full-body control. In pursuit of this, we contribute by designing an optimal control problem that generates trajectories to solve the buzzwire in a time optimized manner. This is composed of task-space constraints to prevent collisions with the buzzwire obstacle, the physical limits of the robot, and an objective function to trade-off reducing time and increasing margins from collision. The formulation can be applied to a very general set of wire shapes and the objective and task constraints can be adapted to other hardware configurations. We evaluate this formulation using the arm of a REEM-C humanoid robot and provide an analysis of how the generated trajectories perform both in simulation and on hardware.