Collision Detection in Legged Locomotion using Supervised Learning

Doshi, Fenil; Brunskill, Emma; Shkolnik, Alexander; Kollar, Thomas; Rohanimanesh, Khashayar; Tedrake, Russ; Roy, Nicholas

doi:10.1109/iros.2007.4399538

Cited by 11 publications

(11 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Machine learning techniques have been used for collision detection [Doshi et al 2007;Pan et al 2011]. However, these techniques cannot be used for PD computation directly.…”

Section: Related Workmentioning

confidence: 99%

Efficient penetration depth approximation using active learning

Pan¹,

Zhang²,

Manocha³

2013

TOG

View full text Add to dashboard Cite

Carolina at Chapel Hill (a) (b) (c) (d) (e) (f) (g) (h) Figure 1: Our algorithm computes a global penetration depth between overlapping non-convex and non-manifold objects. (a) Dynamic simulation of angry bird characters falling into a complex chute in Box2D physics engine; (b) rainfall of 1, 000 rings in Bullet physics engine; (c) a star and a spoon; (d) a spoon and a cup; (e) multiple contacts between upper and lower teeth (each has more than 40, 000 triangles). (f-h) benchmarks consisting of complex models (bunny, dragon and Buddha models have 70K, 230K and 1M triangles, respectively). Our PD algorithm takes less than 0.1 ∼ 2 milliseconds, with less than 2-3% relative error, for each pair of overlapping objects.

show abstract

“…Machine learning techniques have been used for collision detection [Doshi et al 2007;Pan et al 2011]. However, these techniques cannot be used for PD computation directly.…”

Section: Related Workmentioning

confidence: 99%

Efficient penetration depth approximation using active learning

Pan¹,

Zhang²,

Manocha³

2013

TOG

View full text Add to dashboard Cite

show abstract

“…A work presented in [4] proposes an approach that enables a quadruped robot -the LittleDog -to select collision free foot placement positions on rough terrain. After selecting a set of random possible choices, the robot uses a classifier to discern those that are collision free from those that are not.…”

Section: Learning To Classify Possible Foot Positionsmentioning

confidence: 99%

“…The training is executed offline but after it is complete, the robot is aware of its own limitation and is able to choose the safer foot placements. This approach ( [4]) uses machine learning to enhance a controller to achieve adaptable locomotion in rough terrain. Despite the offline learning, it is able to perform adaptable and autonomous locomotion in harsh conditions.…”

Section: Learning To Classify Possible Foot Positionsmentioning

confidence: 99%

Adaptive Quadruped Locomotion: Learning to Detect and Avoid an Obstacle

Matos

Santos

2012

From Animals to Animats 12

View full text Add to dashboard Cite

Autonomy and adaptability are key features in the design and construction of a robotic system capable of carrying out tasks in an unstructured and not predefined environment. Such features are generally observed in animals, biological systems that usually serve as an inspiration models to the design of robotic systems. The autonomy and adaptability of these biological systems partially arises from their ability to learn. Animals learn to move and control their own body when young, they learn to survive, to hunt and avoid undesirable situations, from their progenitors. There has been an increasing interest in defining a way to endow these abilities into the design and creation of robotic systems. This dissertation proposes a mechanism that seeks to create a learning module to a quadruped robot controller that enables it to both, detect and avoid an obstacle in its path. The detection is based on a Forward Internal Model (FIM) trained online to create expectations about the robot's perceptive information. This information is acquired by a set of range sensors that scan the ground in front of the robot in order to detect the obstacle. In order to avoid stepping on the obstacle, the obstacle detections are used to create a map of responses that will change the locomotion according to what is necessary. The map is built and tuned every time the robot fails to step over the obstacle and defines how the robot should act to avoid these situations in the future. Both learning tasks are carried out online and kept active after the robot has learned, enabling the robot to adapt to possible new situations. The proposed architecture was inspired on [14, 17], but applied here to a quadruped robot with different sensors and specific sensor configuration. Also, the mechanism is coupled with the robot's locomotion generator based in Central Pattern Generators (CPG)s presented in [22]. In order to achieve its goal, the controller send commands to the CPG so that the necessary changes in the locomotion are applied. Results showed the success in both learning tasks. The robot was able to detect the obstacle, and change its locomotion with the acquired information at collision time.

show abstract

“…Researchers are beginning to study creatures and their locomotion manners to design robots with more efficient locomotion capabilities. For example, the bipedal robot in [9] based on studying human locomotion, the quadruped walking robot Little Dog in [10] based on studying fourlegged animals, and the hexapod robot RHex in [11] based on studying six-legged insects. In [12], a flying robot is used to deploy and repair a sensor network.…”

Section: Introductionmentioning

confidence: 99%

A Wireless Sensor Network System with a Jumping Node for Unfriendly Environments

Zhang

Song

Qiao

et al. 2012

International Journal of Distributed Sensor Networks

View full text Add to dashboard Cite

Mobile robots have been adopted to repair failed wireless sensor network systems for node damage, battery exhaustion, or obstacles. But most of the robots use wheeled locomotion manner, which does not work well or even fails when confronted with obstacles in uneven terrains. To solve this problem, this paper presents the design of a jumping robot to serve as a robotic node for wireless sensor networks. The robot can jump up to or over obstacles to repair the broken network connections. The robot senses its posture angle by using an acceleration sensor and self-rights automatically by using a pole leg after falling down on the ground. The robot also can steer and adjust its take-off angle by the pole leg. A network monitoring system with the proposed robot is built to test its basic locomotion capabilities and the network repair function. Experimental results show that the robot can jump about 90 cm in height and traverse 50 cm far at a take-off angle of 75 degrees. The robot can repair the network by jumping up to a 10 cm high platform. The proposed system with a jumping node can provide powerful support for applications in unfriendly environments.

show abstract

Collision Detection in Legged Locomotion using Supervised Learning

Cited by 11 publications

References 12 publications

Efficient penetration depth approximation using active learning

Efficient penetration depth approximation using active learning

Adaptive Quadruped Locomotion: Learning to Detect and Avoid an Obstacle

A Wireless Sensor Network System with a Jumping Node for Unfriendly Environments

Contact Info

Product

Resources

About