4 | {% if page.title == "Home" %} 5 | {{ site.title }} | {{ site.description }} 6 | {% else %} 7 | {{ page.title }} | {{ site.title }} 8

ABSTRACT

49 | Legged locomotion can extend the operational domain of robots to some of the most challenging environments 50 | on Earth. However, conventional controllers for legged locomotion are based on elaborate state machines that 51 | explicitly trigger the execution of motion primitives and reflexes. These designs have increased in complexity but 52 | fallen short of the generality and robustness of animal locomotion. Here, we present a robust controller for blind 53 | quadrupedal locomotion in challenging natural environments. Our approach incorporates proprioceptive feedback 54 | in locomotion control and demonstrates zero-shot generalization from simulation to natural environments. The 55 | controller is trained by reinforcement learning in simulation. The controller is driven by a neural network policy 56 | that acts on a stream of proprioceptive signals. The controller retains its robustness under conditions that were 57 | never encountered during training: deformable terrains such as mud and snow, dynamic footholds such as rubble, 58 | and overground impediments such as thick vegetation and gushing water. The presented work indicates that 59 | robust locomotion in natural environments can be achieved by training in simple domains. 60 |

65 | Summary video 66 |

69 | 72 |

76 |

77 |

Highlights

78 |

79 | Zero-shot generalization to unknown environments
80 |

81 | 82 |

83 | Figure 1. A number of specific deployments 84 | (paper Figure 2) 85 |

86 |

89 | 90 | 91 |

92 | The presented controller has been deployed in diverse natural environments. 93 | These include steep mountain trails, creeks with running water, mud, thick vegetation, loose rubble, 94 | snow-covered 95 | hills, and a damp forest. 96 | A number of specific scenarios are further highlighted in Fig. 1A-F. These environments have 97 | characteristics that the policy does not experience during training. The terrains can deform and crumble, with 98 | significant variation of material properties over the surface. The robot's legs are subjected to frequent 99 | disturbances due to vegetation, rubble, and sticky mud. 100 | Existing terrain estimation pipelines that use cameras or LiDAR fail in environments 101 | with 102 | snow (Fig. 1A), water (Fig. 1C), 103 | or dense vegetation (Fig. 1F). 104 | Our controller does not rely on exteroception and is immune to such failure. 105 | The controller learns omnidirectional locomotion based on a history of proprioceptive observations and 106 | is robust in zero-shot deployment on terrains with characteristics that were never experienced during training. 107 |

108 | 109 |

110 | Our controller was used by the Cerberus team for the DARPA Subterranean Challenge Urban Circuit 111 | (Fig. 1G). 112 | It replaced a model-based controller that had been employed used by the team in the past. 113 | The objective of the competition is to develop robotic systems that rapidly map, navigate, 114 | and search complex underground environments, including tunnels, urban underground, and cave 115 | networks. The human operators are not allowed to assist the robots during the competition physically; 116 | only teleoperation is allowed. Accordingly, the locomotion controller needs to perform without failure 117 | over extended mission durations. 118 | The presented controller drove two ANYmal-B robots in four missions of 60 minutes. The 119 | controller exhibited a zero failure rate throughout the competition. A steep staircase that was traversed 120 | by one of the robots during the competition is shown in Fig. 1G. 121 |

122 | 123 |

124 | 125 |

126 | Foot-trapping reflex
127 |

128 | 129 |

130 | Movie 1. Step experiment 131 | (paper Movie S3) 132 |

133 |

134 | 137 |

138 | 139 |

140 | The learned controller manifests a foot-trapping reflex, as shown in Movie 1. 141 | The policy identifies the trapping of the foot purely from proprioceptive observations and lifts the foot over 142 | the obstacle. 143 | Such reflexes were not specified in any way during training: they developed adaptively. 144 | This distinguishes the presented approach from conventional controller design methods, 145 | which explicitly build in such reflexes and orchestrate their execution by a higher-level state machine. 146 |

147 | 148 |

149 | 150 |

151 | Robustness to Model-mismatch
152 |

153 | 154 |

155 | Movie 2. Payload experiment 156 | (paper Movie S4) 157 |

158 |

159 | 162 |

163 | 164 |

165 | We tested the controllers in the presence of substantial model mismatch. 166 | We attached a 10 kg payload. 167 | This payload is 22.7 % of the total weight of the robot, and was never simulated during training. 168 | As shown in Movie 2, the presented controller can still traverse steps up to 13.4 cm 169 | despite the model mismatch. 170 | The baseline is incapable of traversing any steps under any command speed with the payload. 171 |

172 | 173 |

174 | 175 |

176 | Robustness to Foot Slippage
177 |

178 |

179 | Movie 3. Foot Slippage experiment 180 | (paper Movie S5) 181 |

182 |

183 | 186 | 187 | 188 |

189 | Next we test robustness to foot slippage. 190 | To introduce slippage, we used a moistened whiteboard. 191 | The results are shown in Movie 3. 192 | The baseline quickly loses balance, aggressively swings the legs, and falls. In contrast, the presented 193 | controller adapts to the slippery terrain and successfully locomotes in the commanded direction. 194 |

195 | 196 |

198 | 199 |

202 |

Acknowledgment

203 |

204 | 205 | Author contributions: J.L. formulated 206 | the main idea of the training and control methods, implemented the controller, set up the 207 | simulation, and trained control policies. J.L. performed the indoor experiments. J.H. 208 | contributed in setting up the simulation. J.L. and L.W. performed outdoor experiments 209 | together. J.L., J.H., L.W., M.H., and V.K. refined ideas, contributed in the experiment design, 210 | and analyzed the data. 211 |

212 |

213 | Funding: The project was funded, in part, by the Intel 214 | Network on Intelligent Systems, the Swiss National Science Foundation (SNF) through the 215 | National Centre of Competence in Research Robotics, the European Research Council (ERC) 216 | under the European Union’s Horizon 2020 research and innovation programme grant 217 | agreement no 852044 and no 780883. The work has been conducted as part of ANYmal 218 | Research, a community to advance legged robotics. 219 | 220 | 221 | 222 | 223 |

224 |

5 | 6 | {{ site.title | xml_escape }} 7 | {% if site.description %}{{ site.description | xml_escape }}{% endif %} 8 | {{ site.github.url }} 9 | 10 | {% for post in site.posts limit:10 %} 11 | 12 | {{ post.title | xml_escape }} 13 | {% if post.author %} 14 | {{ post.author | xml_escape }} 15 | {% endif %} 16 | {% if post.excerpt %} 17 | {{ post.excerpt | xml_escape }} 18 | {% else %} 19 | {{ post.content | xml_escape }} 20 | {% endif %} 21 | {{ post.date | date_to_rfc822 }} 22 | {{ site.url }}{{ post.url }} 23 | {{ site.url }}{{ post.url }} 24 | 25 | {% endfor %} 26 | 27 |

404: Page not found

7 | {{ page.title }} 8 |

7 | {{ page.title }} 8 |

Paper links

ABSTRACT

Highlights

Zero-shot generalization to unknown environments

Foot-trapping reflex

Robustness to Model-mismatch

Robustness to Foot Slippage

Acknowledgment