Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Towards Autonomous Micromobility through Scalable Urban Simulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces URBAN-SIM, a GPU-parallel urban simulation platform with procedural scene generation, interactive pedestrian dynamics, and asynchronous scene sampling, plus URBAN-BENCH (eight locomotion, navigation, and traverse tasks) benchmarking four robots with PPO.
Problem
Autonomous micromobility (delivery robots, electric wheelchairs) in complex urban public spaces needs embodied agents trained at scale in diverse, realistic urban scenes, but existing robot learning platforms either lack contextual urban environments or lack GPU-parallel training performance.
Approach
Combine hierarchical procedural urban scene generation, GPU-based (JAX) ORCA interactive pedestrian dynamics, and an asynchronous scene sampling scheme that trains parallel environments on many distinct urban scenes simultaneously, then define a benchmark suite spanning locomotion, navigation, and long-horizon traverse tasks.
Key results
- Asynchronous scene sampling yields over 26.3% relative improvement over synchronous sampling at the same training steps, with 1,800-2,600 fps on a single GPU.
- NavClear success rate reaches 97.60% for the wheeled robot, while the quadruped has the lowest collision rates (0.08 in NavStatic, 0.13 in NavDynamic).
- Scaling training scenes from 1 to 1,024 raises NavStatic success rate from 5.1% to 83.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use asynchronous rather than synchronous scene sampling with increasing numbers of unique training scenes | URBAN-SIM on NavStatic · NavStatic | Success Rate | Helped |
| Increase training scenes from 1 to 1,024 | URBAN-SIM on NavStatic · NavStatic | Success Rate: 5.1% to 83.2% | Helped |
| Use wheeled rather than quadruped robots | URBAN-BENCH · NavClear | Success Rate (%): 90.29 ± 3.25 -> 97.60 ± 0.92 | Helped |
| Use a quadruped rather than a wheeled robot | URBAN-BENCH · NavStatic and NavDynamic | Collision: 0.31 ± 0.09 -> 0.08 ± 0.02; 0.35 ± 0.04 -> 0.13 ± 0.02 | Helped |
Surprising findings
- Wheeled robots prefer detouring over even surfaces to avoid stuck risk, rather than taking shorter paths.
- The wheeled robot achieves the highest success in NavClear but drops sharply in NavStatic and NavDynamic, while the humanoid performs best in complex obstacle scenarios.
Limitations
- Urban traverse evaluated only with a quadruped robot and the human-AI shared approach is a pilot study
- Sim-to-real transfer not yet demonstrated (planned as future work)
Experimental setup
Datasets: Urban-Nav Urban-Loc Urban-Tra-Standard Urban-Tra-City
Benchmarks: URBAN-BENCH
Metrics: Balance X-displacement Time to Fall Smoothness Success Rate Route Completion On Walkable Region SPL Collision Human cost
Baselines: synchronous scene sampling (IsaacLab-style)
Builds on: Isaac Gym IsaacLab ORCA PPO Waymax
Closest papers by meaning
- Synthesizing Photorealistic and Dynamic Urban Environments for Multimodal Robot Navigation and Collaboration (NeurIPS 2025 · similarity 0.89)
- CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos (CVPR 2025 · similarity 0.85)
- SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model (CVPR 2025 · similarity 0.85)
- Learning to Drive in New Cities Without Human Demonstrations (NeurIPS 2026 · similarity 0.84)
- Learning to Navigate Efficiently and Precisely in Real Environments (CVPR 2024 · similarity 0.84)
- Vid2Sim: Realistic and Interactive Simulation from Video for Urban Navigation (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.