Environmentally Adaptive Control Including Variance Minimization Using Stochastic Predictive Network with Parametric Bias: Application to Mobile Robots

TL;DR

Addressing robots with stochastic dynamics and partial observability, the paper introduces SPNPB, a stochastic predictive network with a parametric bias that embeds environmental information and outputs a state prediction $\hat{\bm{s}}_{t+1}$ with uncertainty $\hat{\bm{v}}_{t+1}$. The model is trained by maximum likelihood to jointly optimize network weights $W$ and per-environment biases $\bm{p}_{k}$, while online adaptation updates $\bm{p}$ to track environment changes. Control is formulated to minimize both state tracking error and motion variance via $L_{control}= \|\bm{s}^{ref}_{seq}-\hat{\bm{s}}_{seq}\|_{2} + C_{variance} \|\hat{\bm{v}}_{seq}\|_{2} \bm{u}^{opt}_{seq}$, solved with backpropagation to obtain $\bm{u}^{opt}_{seq}$. The approach is validated in simulation and on Fetch, demonstrating environmental recognition through $\bm{p}$ and reduced instability under stochastic friction and surface changes, highlighting potential for adaptive control of flexible and soft robots.

Abstract

In this study, we propose a predictive model composed of a recurrent neural network including parametric bias and stochastic elements, and an environmentally adaptive robot control method including variance minimization using the model. Robots which have flexible bodies or whose states can only be partially observed are difficult to modelize, and their predictive models often have stochastic behaviors. In addition, the physical state of the robot and the surrounding environment change sequentially, and so the predictive model can change online. Therefore, in this study, we construct a learning-based stochastic predictive model implemented in a neural network embedded with such information from the experience of the robot, and develop a control method for the robot to avoid unstable motion with large variance while adapting to the current environment. This method is verified through a mobile robot in simulation and to the actual robot Fetch.

Figures (8)

• Figure 1: The motion of the robot Fetch is stochastic depending on the angles of the casters and the friction of the floor.
• Figure 2: The overall system of using stochastic predictive network with parametric bias (SPNPB).
• Figure 3: Two types of floor materials, Room and Corridor, for Fetch experiment.
• Figure 4: Training experiment in simulation: parametric biases $\bm{p}_{k}$ trained using the collected data and the trajectory of $\bm{p}$ updated by online learning.
• Figure 5: State estimation experiment in simulation: transition of the target velocity $\bm{w}^{ref}$, the predicted velocity $\bm{w}^{pred}$ from SPNPB, and the predicted standard deviation $\bm{\sigma}$ of $\bm{w}^{pred}$.