Research

MAGIC-VFM - Meta-learning Adaptation for Ground Interaction Control with Visual Foundation Models

Elena Sorina Lupu*, Fengze Xie*, James A. Preiss, Matthew Anderson, Jedidiah Alindogan, Soon-Jo Chung

Planning and control of ground vehicles are challenging because of complex dynamic interactions with the terrain.

Therefore, accurate modeling of terrain interaction forces is important to optimize their driving performance.

We present an offline meta-learning algorithm to construct a rapidly-tunable model of residual dynamics and disturbance using both visual foundation models and vehicle states.

This model is then integrated with composite adaptive control for the control matrix to adapt to changes in both the terrain and vehicle dynamics conditions in real time.

Online Policy Optimization in Unknown Nonlinear Systems

Yiheng Lin, James A. Preiss, Fengze Xie, Emile Anand, Soon-Jo Chung, Yisong Yue, Adam Wierman

[paper]

We study online policy optimization in nonlinear time-varying dynamical systems where the true dynamical models are unknown to the controller. This problem is challenging because unlike in linear systems, the controller cannot obtain globally accurate estimations of the ground-truth dynamics using local exploration. We propose a meta-framework that combines a general online policy optimization algorithm with a general online estimator of the dynamical system’s model parameters.

Hierarchical Meta-learning-based Adaptive Controller

Fengze Xie, Guanya Shi, Michael O'Connell, Yisong Yue, Soon-Jo Chung

[paper], [website]

This paper presents the Hierarchical Meta-learning-based Adaptive Controller (HMAC) to learn and adapt to such multi-source disturbances. 

HMAC handles both manageable and latent disturbances with hierarchical iterative learning and smoothed streaming meta-learning. 

Improved Autonomy with Rapidly Learned Dynamics from Adaptive Control

John Lathrop, James Preiss, Elena-Sorina Lupu, Fengze Xie, Soon-Jo Chung

We study the benefits of incorporating a fast-adapting online dynamics model into the robotic autonomy pipeline.

A composite adaptive control law updates the parameters of a residual dynamics model, simultaneously minimizing the disturbance prediction error and the tracking error of the planned trajectory with exponential convergence.

Breaking the strict separation between planning and control, we incorporate online dynamics updates in our motion planner.

Joint-Space Multi-Robot Motion Planning with Learned Decentralized Heuristics

Fengze Xie, Marcus Dominguez-Kuhne, Benjamin Riviere, Jialin Song, Wolfgang Hönig, Soon-Jo Chung, Yisong Yue

[paper]

In this paper, we present a method of multi-robot motion planning by biasing centralized, sampling-based tree search with decentralized, data-driven steer and distance heuristics.