Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Sim2Dust: Mastering Dynamic Waypoint Tracking on Granular Media

Created by
  • Haebom

Author

Andrej Orsula, Matthieu Geist, Miguel Olivares-Mendez, Carol Martinez

Outline

This paper presents a sim-to-real framework for robust autonomous navigation on unstructured terrain on remote planetary surfaces. Considering the dynamics of wheel interactions with complex particle media, we train a reinforcement learning agent through massively parallel simulations in a procedurally generated environment with diverse physical properties. The trained policy is then zero-shot transferred to a real wheeled rover in a lunar-like environment. We compare and analyze several reinforcement learning algorithms and action smoothing filters to identify the most effective combination for real-world deployment. We experimentally demonstrate that agents trained through procedural diversity outperform agents trained in static scenarios in zero-shot performance. We also analyze the tradeoffs of fine-tuning using high-fidelity particle physics, demonstrating that it offers a marginal benefit in improving low-speed accuracy but at a significant computational cost. This establishes a validated workflow for building robust learning-based navigation systems, a significant step toward the application of autonomous robots to space exploration.

Takeaways, Limitations

•
Takeaways:
◦
We propose a deep-to-real transfer learning method based on reinforcement learning that utilizes procedural diversity to improve autonomous driving performance in unstructured terrain.
◦
Through comparative analysis of various reinforcement learning algorithms and action smoothing filters, we present a combination suitable for real-world environments.
◦
We analyze the trade-off between the utility of high-fidelity simulation and computational cost and propose a practical simulation strategy.
◦
The presented framework is expected to make a significant contribution to the development of autonomous driving for future space exploration robots.
•
Limitations:
◦
The effectiveness of high-fidelity particle physics-based fine-tuning is limited. It offers only a marginal benefit in improving low-speed accuracy and is computationally expensive.
◦
Since these are experimental results in a lunar-like environment, further verification is required for generalization performance in actual planetary surface environments.
◦
Because the types of reinforcement learning algorithms and action smoothing filters used are limited, further research into more diverse methodologies is needed.
👍
Made with Slashpage