Saeed Saeedvand

Self-balancing wheeled robots present unique challenges for mobile manipulation due to continuous variations in height, orientation, and tilt caused by the robot’s self-balancing dynamics. In this paper, we present a PPO-based framework with a five-stage dense reward function that shapes the policy to reach, grasp, lift, transport, and accurately place objects on our self-balancing two-wheeled robot under a generalized multiobject training setting, where all three objects (Mug, Drill, and Dumbbell) are trained simultaneously in a single policy. The fivestage reward guides exploration in this high-dimensional floatingbase manipulation task without requiring explicit kinematic programming. Our policy achieves a success rate of 69.78% on the Drill task and sub-centimeter placement accuracy of 9.52 mm. The experiments also reveal the spontaneous emergence of prealignment behavior, where the robot learns to reorient asymmetric objects into graspable poses without explicit instruction, suggesting that the reward structure encourages adaptive manipulation strategies beyond what was explicitly programmed.

Shi-Han Wang, Hanjaya Mandala, Saeed Saeedvand, Jacky Baltes Verified by AVIS CertificateFIRA World Summit 2026Submitted 17 Sept 2026

This paper presents an empirical study evaluating the Multi- Agent Proximal Policy Optimization (MAPPO) algorithm within the Isaac Gym framework. We explore the cooperative behaviors of hu- manoid robots in a multi-agent system tasked with collaboratively mov- ing objects to specified locations. Key performance metrics, such as av- erage and best rewards, which were 1245.87 and 1324.31 respectively, are detailed to illustrate the effectiveness of cooperative strategies. Our findings emphasize how these humanoid agents adapt their strategies for transporting a table by dynamically adjusting their vertical vectors and maintaining an optimal carrying height. The study highlights the agents’ ability to refine and enhance their cooperative maneuvers, facil- itated by the tailored reward functions provided by the MAPPO algo- rithm, demonstrating significant improvements in task execution within a cooperative multi-agent context.

ILHAM AKBAR, Saeed Saeedvand Verified by AVIS CertificateFIRA World Summit 2024Submitted 5 Jul 2024