创新背景

目前，大多数训练有腿机器人行走和导航的方法要么依赖本体感觉，要么依赖视觉，但不能同时依赖两者。在某种情况下，这就像训练一个盲人机器人通过触摸和感觉地面来走路。而在另一个系统中，机器人仅根据视觉来规划腿部运动。

创新过程

由加州大学圣地亚哥分校领导的一个团队开发了一种新的算法系统，使四条腿的机器人能够在具有挑战性的地形上行走和奔跑，同时避开静态和移动障碍物。

在测试中，该系统引导机器人自主快速地在沙质表面、砾石、草地和布满树枝和落叶的崎岖土丘上机动，而不会撞到杆子、树木、灌木、巨石、长凳或人。机器人还可以在繁忙的办公空间中穿行，而不会撞到盒子、桌子或椅子。

该系统将机器人的视觉与本体感觉（本体感觉包括机器人的运动感、方向感、速度感、位置感和触感）结合在一起，为有腿机器人提供了更多的功能。在这项工作中，研究人员将本体感觉与计算机视觉结合起来，使有腿的机器人能够在各种具有挑战性的环境中高效平稳地移动，同时避开障碍，而不仅仅是定义明确的环境。

研究团队开发的系统使用了一套特殊的算法，将机器人头上的深度相机拍摄的实时图像数据与机器人腿上传感器的数据融合在一起。问题是，在实际操作过程中，从相机接收图像有时会有轻微的延迟，所以来自两种不同传感模式的数据并不总是同时到达。

该团队的解决方案是通过随机化两组输入来模拟这种不匹配——研究人员将这种技术称为多模态延迟随机化。然后使用融合和随机输入以端到端方式训练强化学习策略。这种方法帮助机器人在导航过程中快速做出决定，并提前预测环境的变化，因此它可以在不同类型的地形上更快地移动和躲避障碍，而无需操作员的帮助。

创新关键点

该系统将机器人的视觉与本体感觉（本体感觉包括机器人的运动感、方向感、速度感、位置感和触感）结合在一起，为有腿机器人提供了更多的功能。

创新价值

这项工作使研究人员距离制造能够执行搜索和救援任务或在对人类来说过于危险或困难的地方收集信息的机器人又近了一步。

The new algorithm system enables the quadruped robot to move flexibly on rough terrain

A team led by the University of California, San Diego, has developed a new algorithm system that enables four-legged robots to walk and run over challenging terrain while avoiding static and moving obstacles.

In tests, the system guided the robot to maneuver autonomously and quickly over sandy surfaces, gravel, grass and rugged mounds strewn with branches and fallen leaves without bumping into poles, trees, shrubs, boulders, benches or people. Robots can also navigate busy office Spaces without bumping into boxes, desks or chairs.

The system combines the robot's vision with proprioception, which includes the robot's sense of motion, direction, speed, position and touch, to provide more functions for legged robots. In this work, the researchers combined proprioception with computer vision to enable legged robots to move efficiently and smoothly through a variety of challenging environments while avoiding obstacles, not just well-defined ones.

The system developed by the research team uses a special set of algorithms to fuse real-time image data from a depth camera on the robot's head with data from sensors on its legs. The problem is that in practice, there is sometimes a slight delay in receiving images from the camera, so the data from the two different sensing modes don't always arrive at the same time.

The team's solution was to simulate this mismatch by randomizing the two sets of inputs - a technique the researchers call multimodal delayed randomization. Reinforcement learning strategies are then trained in an end-to-end manner using fusion and random inputs. This approach helps the robot make quick decisions during navigation and anticipate changes in the environment in advance, so it can move and avoid obstacles more quickly over different types of terrain without the help of an operator.