Multi-Agent Reinforcement Learning (MARL) plays a crucial role in robotic coordination and control, yet existing simulation environments often lack the fidelity and scalability needed for real-world applications. In this work, we extend Isaac Lab to support efficient training of both homogeneous and heterogeneous multi-agent robotic policies in high-fidelity physics simulations. Our contributions include the development of diverse MARL environments tailored for robotic coordination, integration of Heterogeneous Agent Reinforcement Learning (HARL) algorithms, and a scalable GPU-accelerated framework optimized for large-scale training. We evaluate our framework using two state-of-the-art MARL algorithms—Multi-Agent Reinforcement Learning with Proximal Policy Optimization (MAPPO) and Heterogeneous Agent Reinforcement Learning with Proximal Policy Optimization (HAPPO)—across several robotic tasks. Our results confirm the feasibility of training heterogeneous agents in high-fidelity environments while maintaining the scalability and performance benefits of Isaac Lab. By advancing realistic multi-agent learning at scale, our work lays a foundation for more MARL research in physics-driven robotics. The source code and demonstration videos are available at https://some45bucks.github.io/IsaacLab-HARL/.

This paper presents POSEIDON-SAT, a novel dataset augmentation method designed to enhance the detection of fishing vessels using optical remote sensing technologies. Illegal fishing poses a significant threat to conservation and economic fishing zones, and its detection is often hindered by tactics such as the disabling or manipulation of Automatic Identification System (AIS) transponders. While convolutional neural networks (CNNs) have shown promise in ship detection from optical imagery, the fine-grained classification of fishing vessels is limited by the scarcity of detailed datasets, as these vessels are often underrepresented in existing databases. POSEIDON-SAT addresses this gap by augmenting datasets with synthesized fishing vessel instances, improving the performance of ship detection models, particularly in low-resource scenarios. This approach is tailored for use on low-power, edge computing platforms aboard small satellites, such as CubeSats, where computational resources are highly constrained. By comparing POSEIDON-SAT to traditional class-weighting techniques, we evaluate its impact on lightweight YOLO models optimized for real-time detection aboard such satellites. Our experimental results demonstrate that POSEIDON-SAT significantly improves detection accuracy while reducing false positives, making it an effective tool for enhancing the capabilities of remote sensing platforms in monitoring illegal fishing. This method holds promise for addressing the global challenge of illegal fishing through scalable, efficient satellite-based monitoring systems.

Autonomous navigation through unstructured terrains has been most effectively demonstrated by animals, who utilize a large set of locomotive styles to move through their native habitats. While legged robots have recently demonstrated several of these locomotion modalities (such as walking, running, jumping, and climbing vertical walls), motion planners have yet to be able to leverage these unique mobility characteristics. In this article, we outline some of the specific motion planning challenges faced when attempting to plan for legged systems with dynamic gaits, with specific instances of these demonstrated by the dynamic climbing platform TAILS. Using a unique implementation of sampling-based model predictive optimization, we demonstrate the ability to motion plan around obstacles on vertical walls and experimentally demonstrate this on TAILS by navigating through traditionally difficult narrow gap problems.

Energy-efficient navigation is an important technology for mobile robots because of its potential to increase the operation time of the robot. In particular, when coupled with a dynamic legged quadruped, the need for energy savings is made more apparent as payloads are limited. Due to the complexity in modeling motion and power models of these robots, a new approach is necessary to effectively motion plan for these complex robots. We accomplish this by using Sampling-Based Model Predictive optimization (SBMPO) which was extended for use on the LLAMA quadrupedal platform in simulation. SBMPO allows for direct generation of trajectories while using a heuristic-based search to speed up computations. This approach is shown to effectively motion plan while optimizing for energy consumption and maintaining the natural dynamics of the robot in a simulated environment.