Introduction
Autonomous micromobility—encompassing lightweight delivery robots, electric scooters, and automated wheelchairs—presents a unique set of navigation challenges distinct from those of autonomous vehicles. While cars operate on structured roads, micromobility devices must navigate the "wild west" of urban public spaces, including crowded sidewalks, steep ramps, stairs, and unpredictable pedestrian traffic.
To address these challenges, we introduce URBAN-SIM, a high-performance simulation platform designed for massive-scale reinforcement learning. URBAN-SIM bridges the gap between high-speed but visually simple simulators and detailed but slow platforms. By achieving over 1,800 FPS on a single GPU, it enables the training of agents across thousands of diverse, realistic urban scenes simultaneously.
Methodology
URBAN-SIM is built on NVIDIA Omniverse and PhysX 5, utilizing three core innovations to ensure scalability and realism:
- Hierarchical Urban Generation: The platform employs a procedural pipeline that constructs cities from the top down. It begins with road block layouts, divides the ground into functional zones such as sidewalks and parks, and finally generates micro-terrains (e.g., stairs, slopes) and populates the scene with 3D objects.
- Interactive Dynamics Generation: To simulate realistic pedestrian and cyclist behavior, path-planning logic (ORCA) is moved entirely onto the GPU using JAX. This allows thousands of dynamic agents to interact with the robot in real-time without the typical CPU-GPU bottleneck.
- Asynchronous Scene Sampling: Unlike standard simulators that train batches in identical environments, URBAN-SIM uses an asynchronous scheme. Each parallel environment samples different assets and layouts from a cache, ensuring the robot encounters a vast variety of edge cases during every training step.
Evaluation: URBAN-BENCH
We evaluate performance using URBAN-BENCH, a suite designed to test three core competencies:
- Urban Locomotion: Movement across flat ground, slopes, stairs, and rough terrain.
- Urban Navigation: Pathfinding through clear areas, static obstacle avoidance, and dynamic pedestrian avoidance.
- Urban Traverse: A long-range task requiring robots to navigate over 1 km through a complex city.
The benchmark tests four distinct robot embodiments: a wheeled delivery bot (COCO), a quadruped (Unitree Go2), a wheeled-legged hybrid (B2-W), and a humanoid (Unitree G1).
Results and Insights
Our experiments highlight the critical role of environmental diversity. Training on a single environment led to poor generalization, whereas asynchronous sampling yielded a 26.3% relative improvement in training efficiency. Furthermore, increasing the number of unique training scenes from 1 to 1,024 caused the success rate on navigation tasks to jump from 5.1% to 83.2%.
Key performance metrics include:
- Throughput: 1,800 to 2,600 FPS with RGBD sensors on a single NVIDIA L40S GPU.
- Embodiment Strategies: Physical form significantly influenced navigation strategies. Humanoids learned to sidestep through narrow gaps, while wheeled robots prioritized finding ramps to bypass stairs.
- Shared Autonomy: For long-horizon tasks like the 1km "Traverse," a "Human-AI Shared Autonomy" model proved most effective, where AI handles routine navigation and humans intervene in high-risk scenarios.
Conclusion and Future Work
URBAN-SIM demonstrates that scalable, diverse simulation is key to advancing autonomous micromobility. Future work will focus on integrating OpenStreetMap data to create "digital twins" of real cities and expanding the benchmark to include manipulation tasks, such as parcel pickup. We also aim to bridge the sim-to-real gap through a dedicated ROS2-based deployment pipeline.
Summary updated