1. Robots that can adapt like animals (Nature cover article)
Antoine Cully, Jeff Clune, Danesh Tarapore, Jean-Baptiste Mouret
The Intelligent Trial and Error Algorithm introduced in the paper ‘Robots that can adapt like animals’ (Nature, 2015): the video shows two different robots that can adapt to a wide variety of injuries in under two minutes. A six-legged robot adapts to keep walking even if two of its legs are broken, and a robotic arm learns how to correctly place an object even with several broken motors.
Full citation: Cully A, Clune J, Tarapore DT, Mouret J-B. Robots that can adapt like animals. Nature, 2015. 521.7553, (cover article).
Source code: https://github.com/resibots/cully_2015_nature
Arxiv (free): http://arxiv.org/abs/1407.3501
2. Analysis of human push recovery motions during walking
R. Malin Schemschat, Debora Clever, Martin L. Felis, Katja Mombaur
As dangers that can disturb a human motion lurk everywhere, the analysis of human
push recovery motions is crucial for the development of stable humanoids. In our work we focus on the analysis of human walking motions that are disturbed by pushes from the back with various strength at different hight at the spine. The analysis includes motion capture experiments, including the recording of the force of the disturbance, as well as two simulation approaches. In both approaches, we use a dynamical rigid multi-body model of a human with 14 segments connected by 13 joints and define an optimal control problem that is solved by the multiple shooting algorithm MUSCOD II. In the first approach, we use the data from motion capture experiments and generate a motion that is as close as possible to the reference data, while fulfilling the constraints of our model. In the second approach, the resulting motion is optimal based on a specific objective function. The motions presented in the video result from a combined objective function, that minimizes the effort and makes the resulting step as periodic as possible.
3. Robust Optimal Control and High-Level Planner Synthesis over Various Terrain Topologies
Ye Zhao and Luis Sentis
This video is composed of three simulation scenarios. We first demonstrate dynamic walking over rough terrains with randomly generated height variations. This behavior validates the versatility of our phase-space planning and control strategy. The lateral foot placement is searched according to a Newton-Raphson algorithm. Walking step transitions are achieved when the phase-space trajectories of two adjacent walking steps intersect with each other. The terrain has concave steps and a 10◦ tilt angle is used for the slope of the steps.
In the second scenario, we apply pushing forces to the lateral and sagittal directions, respectively, both of which cause instantaneous center of mass velocity jumps. For the lateral disturbance, a new lateral foot placement is re-planned according to the searching algorithm. For the sagittal disturbance, the disturbance is quite large such that the robot’s state cannot recover in one single step to its nominal phase-space manifold using the proposed optimal controller. Thus, the foot location re-planning strategy is executed in a combinational manner with the optimal controller.
In the last scenario, the robot is maneuvering within a constrained environment by properly using both leg and arm contacts. A temporal logic based high-level planner is synthesized to make contact decisions according to possibly adversarial environments, including emergent behaviors such as stair crack and human appearance.
4. Real-time Slip Detection and Recovery for a Hexapod Robot
Benjamin Tam, Navinda Kottege
We present a system for real-time slip detection and recovery for an 18 DoF hexapod robot. When executing the 18 dimensional gait cycle, the robot is subject to various disturbances due to joint compliance and foot-ground interactions. Leg slip due to loss of traction is an example. Therefore, the actual gait cycle in configuration space can differ from the planned gait cycle. We call this difference the configuration error δc defined as the Euclidean distance between the planned and actual configurations. The Stabilisation system constantly keeps track of this configuration error and reacts when it exceeds a threshold (Δc). The algorithms also keeps track of touchdown configurations (when all legs are on the ground) considered to be stable states. These configurations are used as intermediate states for recovery with the onset of a slip event. The video shows an experiment where the robot walks on a ‘random slip field’, and the algorithms detects and recovers from slip events.