Deep Robotic Grasping using Generative Residual Convolutional Neural Network

A CNN (GR-ConvNet) model that can generate robust antipodal grasps from RGB-D input images at real-time

Setup

• Clone this repository and install required libraries

git clone https://github.com/Loahit5101/GR-ConvNet-grasping.git
cd robotic-grasping
pip install -r requirements.txt

• Install TensorRT and Torch-TensorRT

Architecture

Dataset

Download and extract the Cornell Grasping dataset and run the following command: Cornell Grasping Dataset

python -m utils.dataset_processing.generate_cornell_depth <Path To Dataset>

Pretrained models

Trained models are available here

Training

python train.py

Testing

python test.py

Model Optimization using TensorRT

• Post-training quantization

python ptq.py

• Benchmarking grasp inference time of optimized and unoptimized models

python trt_benchmark.py

Results

Average Grasp Inference Time

Model	Time (ms)	Accuracy
Baseline	4.59	95.5
FP-32	3.71	94.2
FP-16	1.45	93.16

Post-processing time = 14 ms on average

References

1. Antipodal Robotic Grasping using GR-ConvNet
2. Pytorch-TensorRT Tutorials