Micromobility solutions—such as delivery robots, mobility scooters, and electric wheelchairs—are rapidly transforming short-distance urban travel. Most micromobility solutions still heavily rely on humans to operate them, despite their increasing popularity as eco-friendly, flexible transport options. The lack of automation limits the efficiency of micromobility devices and increases safety risks, particularly in urban environments with many dynamic objects like cyclists and pedestrians.
Autonomous Micromobility: A Need in Urban Spaces
Traditional transportation methods like cars and buses are ideal for long-distance travel but often struggle with last-mile connectivity—the final leg in urban journeys. The micromobility technology fills the gap with lightweight devices, which are low-speed and ideal for urban short trips. True autonomy is still elusive in the micromobility space. Current AI solutions are focused on narrow tasks, such as simple navigation or obstacle avoidance, but fail to tackle complex urban challenges, including uneven terrain and crowds.
The limitations of existing robot learning and simulation platforms
Many robot-training simulation platforms are designed to simulate indoor environments, vehicle-centric roads, and other similar situations, but lack the richness of context found on urban alleyways and sidewalks. In contrast, high-efficiency platforms provide scenes that are not suitable for deep learning when there are obstacles or unpredictable movements of pedestrians. The AI agent’s ability to learn the critical skills required for micromobility is severely limited by this gap.
Introducing URBAN SIM: High Performance Simulation for Micromobility in Urban Areas
In order to address these issues, researchers at the University of California Los Angeles (UCLA) and the University of Washington created URBAN-SIMA high-fidelity, scaleable urban simulation platform specifically designed for research on autonomous micromobility.
Key Features of URBAN SIM
- Hierarchical City Scenes Generation
Procedurally creates infinitely diverse, large-scale urban environments—from street blocks to detailed terrain features—that include sidewalks, ramps, stairs, and uneven surfaces. The layered pipeline creates realistic, varied environments for robot training. - The Interactive Dynamic agent Simulation
Real-time simulation of pedestrians, bicycles, and cars on GPUs. Complex multi-agent interaction that simulates urban dynamics. - Scalable Asynchronous Sampling of Scenes
Allows AI agents to be trained in parallel on hundreds of complex and unique urban scenes using a single graphics card. This dramatically increases training speed, and promotes robust policy learning.
URBAN SIM, based on NVIDIA Omniverse’s PhysX physics engines, combines realistic rendering and precision physics to provide authentic embodied AI.
URBAN-BENCH: Comprehensive Benchmark Suite for Real-World Skills
Created to compliment URBAN SIM, this team has created URBAN-BENCHThe URBAN-BENCH includes: URBAN-BENCH includes:
- Urban Locomotion Tasks: The robot can move on slopes, rough terrain, or flat surfaces.
- Urban Navigation Tasks: Avoiding static obstructions like trash cans and benches, while managing moving obstacles, such as pedestrians or cyclists.
- Urban Traverse Task: This kilometer-scale challenge combines terrain, obstacles, and moving agents to test navigational skills.
A Shared Human-AI Autonomy Approach
Urban-Benches are designed to perform long-distance urban navigation tasks. The human-AI autonomy shared model. This flexible control architecture decomposes the robot’s control system into layers—high-level decision making, mid-level navigation, and low-level locomotion—allowing humans to intervene in complex or risky scenarios while enabling AI to manage routine navigation and movement. In dynamic urban areas, this collaboration helps balance safety with efficiency.
Evaluation of Diverse Robots for Realistic Tasks
URBAN BENCH and URBAN SIM support a variety of robot platforms including humanoid, wheeled and quadrupedal robots. The benchmarks show the strengths and weaknesses of each type of robot in locomotion challenges and navigation problems, which illustrates how generalizable this platform is.
Example:
- The quadruped robot excels in terms of stability and the ability to climb stairs.
- The best performance of wheeled robots is on flat, clear paths.
- This hybrid design allows for a combination of terrain adaptation and wheeled-legged robots.
- Sidestepping allows humanoid robotic robots to navigate narrow urban spaces.
Training efficiency and Scalability
This strategy allows for training on diverse urban environments, and has shown to improve performance by up to 26,3% over traditional synchronous training. Increasing diversity in training environments correlates directly with greater success rates for navigation tasks. This highlights the importance of large-scale simulation for robust, autonomous micromobility.
The conclusion of the article is:
URBAN BENCH & URBAN SIM are crucial steps in enabling safe and efficient autonomous micromobility within complex urban contexts. Through ROS 2 integration, and using sim-to real transfer techniques, future efforts will aim to connect simulation with actual deployment. The platform will also evolve in order to include multi-modal perception, manipulation and control capabilities that are necessary for urban robot applications like parcel delivery or assistive robotics.
By enabling scalable training and benchmarking of embodied AI agents in authentic urban scenarios, this research catalyzes progress in autonomous micromobility—promoting sustainable urban development, enhancing accessibility, and improving safety in public spaces.
Take a look at the Paper You can also find out more about the following: Code. This research is the work of researchers on this project. SUBSCRIBE NOW Subscribe to the AI Newsletter
Sana Hassan is a dual-degree IIT Madras student and consulting intern with Marktechpost. She loves to apply technology and AI in order to solve real-world problems. Sana Hassan, an intern at Marktechpost and dual-degree student at IIT Madras, is passionate about applying technology and AI to real-world challenges.
