PDF Paper ArXiv arXiv Youtube Video Code

AnyCar to Anywhere

Learning Universal Dynamics Model
for Agile Mobility

Wenli Xiao*   Haoru Xue*   Tony Tao   Dvij Kalaria   John M. Dolan   Guanya Shi  
*equal contributions

In Submission

PDF Paper ArXiv arXiv Youtube Video Code

Audio Overview (generated using NotebookLM)

Our Approach

AnyCar System Pipeline. (1) Data Collection: We collect data in 4 different simulations to pre-train the AnyCar model. We separately also collect real-world data for fine-tuning of the model. (2) Model Training: We pre-train the model with the simulation dataset and enhance prediction robustness through masking, adding noise, and attacking the inputs. (3) Deployment: We fine-tune the model with few-shot real-world data to achieve zero-shot transfer across different vehicles, environmental conditions, and state estimation methods.

Deployment In the Wild

Deployment In Uneven Terrain

Large-Scale Pre-training in Diverse Simulation

Few-shot Real-world Fine-tuning (Teleoperation)

Out-of-Domain Adaptation

One More Thing ...

We Open-Source Our 1/16 Car Design for Wheeled Agility Research

BibTeX

@misc{xiao2024anycaranywherelearninguniversal,
      title={AnyCar to Anywhere: Learning Universal Dynamics Model for Agile and Adaptive Mobility}, 
      author={Wenli Xiao and Haoru Xue and Tony Tao and Dvij Kalaria and John M. Dolan and Guanya Shi},
      year={2024},
      eprint={2409.15783},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2409.15783}, 
}