Module Pool¶

DeepClaw maintains a module pool of algorithms and end-to-end robot learning models by integrating some state-of-the-art research results in computer vision and robotics. The codes are placed under deepclaw/modules/.

Computation on server¶

As the running environment for each method is different, DeepClaw adopts concepts from cloud robotics. We put the running environments for end-to-end methods which requires heavy computations in a docker container on the server, and deploy the robot control and basic computations on a user computer.

If you want to add a new module running on server and runing the inferene from a client computer, please refer to the tutorial server and client.

Currently we have two servers: Goldenboy and Serbreeze. Currently they are running Ubuntu16.04 and cuda9.0. We plan to upgrade them to Ubuntu18.04 and cuda10 soon. Each user is assigned to have one GPU card by setting environment varible CUDA_VISIBLE_DEVICES. Please don't change it by yourself. If you need more computation resources, please contact us.

	Goldenby	Serbreeze
Memory	251.8G	125.8 GiB
Processor	Intel® Xeon(R) CPU E5-2698 v4 @ 2.20GHz × 40	Intel® Xeon(R) CPU E5-2650 v4 @ 2.20GHz × 48
GPU	Tesla V100 32G x4	GeForce GTX 1080Ti 12G x4
Storage	7.6TB SSD	240G SSD (/home), 960GB SSD+8TB HD (/media/amax/)
Users	Standard: user-1, user-2, user-3	Standard: student1, student2, student3
IP	10.20.123.35	10.20.73.134

List of modules¶

We list all the algorithms and model available in DeepClaw. The code can be found under deepclaw/modules/ and demos are included under algorithm folders. Please check the following notes before running the demos: - In addition to the requirement of deepclaw, you might need to install extra dependencies for each module. Please refer to the readme.md under each module for more instructions. - You might need to download checkpoint of weight to run the demo. Please refer to the list below for downloading pretrained weights.

Segmentation¶

Method	Object classes	weights
Contour detector	NA	NA

Recognition¶

Method	Object classes	weights
Efficientnet	4 recyclable waste	link extract code: ph2e

Object Detection¶

Method	Object classes	weights
Efficientdet	204 waste classes	link extract code: frra

Grasp Planning¶

Method	output	weights
GraspNet	9 rotation angles	link extract code: 7hy5
DexNet	gras pose with robustness	refer to the readme

Motion Planning¶

Method	output	weights
Predefined waypoints