Micromobility solutions – such as delivery robots, mobility scooters and electric wheelchairs – quickly transform short -distance urban trips. Despite their growing popularity as a flexible and environmental transportal transport, most of the micromobility devices are still strongly based on human control. This dependence limits operational efficiency and raises security problems, especially in complex crowded city environments filled with dynamic obstacles such as pedestrians and cyclists.
The need for autonomous micromobility in urban areas
Traditional transport methods such as cars and buses are ideal for long -distance trips, but often have trouble with the connectivity of the last kilometer – the last stage of urban trips. The micromobility fills this gap by offering light and low speed devices that excel in short urban trips. However, real micromobility autonomy remains elusive: AI's current solutions tend to focus closely on specific tasks such as avoidance of obstacles or simple navigation, not raising the challenges with multiple facets posed by real urban environments which include uneven terrain, stairs and dense crowds.
Limits of existing robots learning and simulation platforms
Most simulation platforms for the training of robots are adapted to interior environments or road networks centered on the vehicle and lack wealth and contextual complexity found in urban sidewalks, places and alleys. Meanwhile, very effective platforms often provide simplified scenes unfit for in -depth learning in environments with various obstacles and unpredictable pedestrian movements. This difference restricts the ability of AI agents to effectively acquire critical skills for autonomous micromobility.
Presentation of urban SIM: high performance simulation for urban micromobility
To meet these challenges, researchers from the University of California in Los Angeles and the University of Washington Urban SimA scalable and high fidelity urban simulation platform designed explicitly for autonomous micromobility research.


Key characteristics of urban SIM:
- Hierarchical urban scene generation
Create urban environments on a large scale infinitely diversified on a large scale – street blocks with detailed characteristics of the terrain – which include sidewalks, ramps, stairs and unequal surfaces. This layer pipeline provides a realistic and varied frame for the training of robots. - Interactive dynamic agent simulation
Simulates pedestrians, cyclists and reactive vehicles in real time on GPUs, allowing complex multi-agent interactions that imitate the true urban dynamic. - Asynchronous scene sampling for scalability
Allows parallel training of AI agents in hundreds of unique and complex urban scenes on a single GPU, considerably increasing training speed and promoting robust political learning.
Built on the omniverse and Physx Physics of Nvidia, Urban-Sim combines a realistic visual rendering with precision physics for authentic training embodied in AI.
Urban Bench: complete reference suite for real world skills
Complete Urban-Sim, the team has created Urban benchA series of tasks and a reference frame that captures essential autonomous micromobility of micromobility based on real urban use scenarios. The urban bench includes:
- Urban locomotion tasks: Crossing flat surfaces, slopes, stairs and rough terrain to ensure a stable and effective robot movement.
- Urban navigation tasks: By browsing clear routes, avoiding static obstacles such as benches and garbage cans and manage dynamic obstacles such as pedestrians and cyclists in motion.
- Urban crossing task: A difficult journey across the kilometers combining land, obstacles and complex dynamic agents, designed to test navigation and decision -making in long horizon.
Shared human shared autonomy approach
For the long -distance urban crossing task, Urban Bench introduced Autonomy model shared by man-a. This flexible control architecture decomposes the robot control system in layers – high level decision -making, intermediate level navigation and low -level locomotion – going to intervene in complex or risky scenarios while allowing AI to manage navigation and routine movement. This collaboration balances the safety and efficiency of dynamic urban environments.
Evaluate various robots in realistic tasks
Urban-Sim and Urban Bench support a wide range of robotic platforms, including rolling, quadrupted, wheel and humanoid robots. The references reveal unique strengths and weaknesses for each type of robot through the challenges of locomotion and navigation, illustrating the generalization of the platform.
For example:
- The quadruped robots excel in stability and crossing the stairs.
- The wheel robots work better on clear and dishes.
- Robots on paws take advantage of their hybrid design for combined adaptability on the grounds.
- Humanoid robots are effectively navigating in narrow and crowded urban areas by bypassing.
Scalability and efficiency of training
The asynchronous scene sampling strategy allows training in various urban scenes, demonstrating up to 26.3% performance improvement compared to synchronous training methods. The increase in the diversity of training environments is directly correlated with higher success rates in navigation tasks, emphasizing the need for large -scale and varied simulation for robust autonomous micromobility.
Conclusion
The urban SIM and the urban bench represent essential stages towards it allowing a safe, efficient and evolving autonomous micromobility in complex urban environments. Future work aims to fill the simulation and the deployment of the real world thanks to ROS 2 integration techniques and Sim-à -Rél transfer. In addition, the platform will evolve to integrate multimodal perception and manipulation capacities necessary for complete urban robot applications such as package delivery and assistance robotics.
By allowing evolutionary training and a comparative analysis of integrity agents embodied in authentic urban scenarios, this research catalyzes progress in autonomous micromobility – promising sustainable urban development, improving accessibility and improving security in public spaces.
Discover the Paper And Code. All the merit of this research goes to researchers in this project. Subscribe now to our newsletter IA
Sana Hassan, consulting trainee at Marktechpost and double -degree student at Iit Madras, is passionate about the application of technology and AI to meet the challenges of the real world. With a great interest in solving practical problems, it brings a new perspective to the intersection of AI and real life solutions.