Humanoid robots are designed to operate in environments built for humans,
making whole-body manipulation, the simultaneous use of arms and legs to
reach, grasp, and move objects, one of the key challenges in modern robotics.
Learning such tasks with reinforcement learning is challenging due to the
high-dimensional action space and complex rewards, especially when all reward
components are introduced simultaneously, as the agent tends to make compromises
between conflicting objectives and converge to a poor local minimum rather than
progressively acquiring individual skills.
The main goal of this work is to learn a policy for stable whole-body manipulation of an object with a humanoid robot, where the robot must learn to perform the task effectively while simultaneously maintaining stability and balance. To address this problem, a progressive approach is implemented, where the task is
divided into three sequential learning phases, each building on the policy of
the previous phase by introducing new reward functions. The Talos humanoid robot
is first trained to maintain balance under external disturbances, then to reach
for an object on a table, and finally to grasp and lift it. Training is performed in the IsaacLab simulation environment using reinforcement learning, with PPO (Proximal Policy Optimization) used for policy optimization.
The results confirm the effectiveness of the progressive approach, as the robot successfully acquired the sequential learning phases, from maintaining balance to performing the manipulation task. In the final phase, the robot successfully grasped and lifted the object, demonstrating the potential of the proposed approach for learning complex whole-body behaviors in humanoid robots.
|
