Repairing Neural Networks for Safety in Robotic Systems using Predictive Models

Abstract

This paper introduces a new method for safety-aware robot learning, focusing on repairing policies using predictive models. Our method combines behavioral cloning with neural network repair in a two-step supervised learning framework. It first learns a policy from expert demonstrations and then applies repair subject to predictive models to enforce safety constraints. The predictive models can encompass various aspects relevant to robot learning applications, such as proprioceptive states and collision likelihood. Our experimental results demonstrate that the learned policy successfully adheres to a predefined set of safety constraints on two applications: mobile robot navigation, and real-world lower-leg prostheses. Additionally, we have shown that our method effectively reduces repeated interaction with the robot, leading to substantial time savings during the learning process.

Figures (9)

• Figure 1: Safety-Aware Repair with Predictive models (SARP). Left: A policy trained for a mobile navigation task (point to goal motion planning in this figure). Right: Policy repair module that adjusts the policy parameters to penalize unsafe behavior, based on a set of safety constraints and the loss of predictive models. Predictive models may include state-action transition model, a model of environment, or a proprioceptive model predicting internal states of the system.
• Figure 2: A simulation of a mobile robot in a hospital scenario. The robot is tasked with getting to different rooms without colliding with the environment.
• Figure 3: Lower-leg prosthesis system. (left) Image depicts the upper limb angle $\alpha_{ul}$, the lower limb angle $\alpha_{ll}$, and the ankle angle $\alpha_{a}$. (middle) The location of pressure sensors $p_1$-$p_{16}$. (right) The robotic lower-limb prosthesis device.
• Figure 4: Navigation in hospital: the collision avoidance of the robot improves by over $93\%$ with SARP shown in (b) compared to the original policy depicted in (a).
• Figure 5: Navigation in hospital: (a)-(b) minimum range sensor values and velocity distributions before and after constraint application. (c) Goal-reaching accuracy and the percentage of safe samples vs. the accuracy of the prediction model.

Theorems & Definitions (1)

• Remark 1