RoboPapers Ep#14 VERTIFORMER: A Data-Efficient Multi-Task Transformer on Vertically Challenging Terrain
Jun 16, 2025
In this enlightening discussion, Xuesu Xiao, an assistant professor of computer science at George Mason University and a roboticist, shares insights on advanced robotic motion planning. He delves into the innovative Vertiformer model designed for navigating difficult terrains, emphasizing the fusion of physics and machine learning. The conversation highlights challenges in estimating flight durations, the importance of data efficiency in transformers, and the integration of real-world testing for enhancing vehicle performance. Get ready to explore the future of robotics!
AI Snips
Chapters
Transcript
Episode notes
6D Motion Planning for Tough Terrain
- Robots can navigate vertically challenging terrain using minimal mechanical complexity by leveraging motion planning and learning techniques.
- Accurately planning in 6D (XYZ, roll, pitch, yaw) avoids vehicle rollovers and getting stuck between obstacles.
Flying RC Car Experiment
- They launched a regular RC car off a ramp to test in-air control using only throttle and steering.
- Many robots broke during these experiments as they tested landing with controlled attitude in air.
Physics Enables In-Air Control
- Controlling the pitch and roll of a car in the air is possible via inertia and gyroscopic effects of spinning wheels.
- Front wheels affect both roll and pitch while rear wheels mainly influence pitch, allowing some in-air attitude control without added hardware.