FPGA-Based Implementation of Stochastic Configuration Network for Robotic Grasping Recognition

Jiaqi Pan, Feng Luan, Yunqi Gao, Yangjie WeiIEEE Access 2020 · OpenReview

Abstract

Precise and rapid grasping recognition based on machine vision is one of the challenging problems for intelligent robots. As a nonlinear network for recognition, the stochastic configuration network (SCN) is considered as a promising method due to its universal approximation capability and fast modelling property. In this paper, SCN with depth image is applied to robotic grasping recognition for the first time. Moreover, to obtain better grasping performance, a novel autoencoder-based SCN (SCN-AE) is proposed to extract the higher level features of the captured image. Finally, a field programmable gate array (FPGA) framework for implementing SCN models is designed and analyzed, which can effectively balance the recognition accuracy and resource utilization. Experimental results on the Cornell grasping dataset proved that the proposed SCN-AE method achieved higher accuracy and better stability in robotic grasping recognition tasks, with an average recognition accuracy of 91.935%.

Introduction

For an intelligent robot, "seeing" an object is only half the battle; the real challenge is grasping recognition—determining exactly where and how to place a gripper to successfully pick something up. While depth cameras (like RGB-D sensors) provide the 3D data needed for this task, processing that data in real-time on a robot’s local hardware is difficult.

Traditional deep learning models often require massive computational power and long training times. Conversely, "random weight networks" like the Extreme Learning Machine (ELM) are fast but can be unstable or inaccurate. Furthermore, most of these models run in software, which can be too slow for the split-second decisions required in dynamic environments. This research aims to bridge this gap by creating a model that is fast to train, highly stable, and optimized for low-power hardware (FPGAs).

Methodology

The proposed solution involves a two-pronged approach: a new algorithm and a custom hardware architecture.

SCN-AE: Stochastic Configuration Network with Autoencoder

Unlike standard neural networks that use backpropagation, a Stochastic Configuration Network (SCN) builds itself by adding hidden nodes one by one. It uses a specific "inequality constraint" to pick random weights that are guaranteed to reduce the model's error, making it more stable than previous random-weight methods.

Architecture of the Stochastic Configuration Network (SCN).

The authors introduced an unsupervised autoencoder layer to the SCN (SCN-AE). This allows the system to extract high-level, abstract features from raw depth images before classifying them as "graspable" or "not graspable."

Flowchart of the proposed SCN-AE algorithm.

FPGA Implementation

To move the model from a desktop computer to a robot's "brain," a custom framework for a Field Programmable Gate Array (FPGA) was designed. Key optimizations include:

  • Fixed-Point Arithmetic: Instead of power-hungry floating-point numbers, a 16-bit fixed-point encoding (1:4:11 format) was utilized.
  • Linearized Activation: The standard "sigmoid" function was replaced with a piecewise linear approximation to save hardware resources while maintaining accuracy.
  • Parallel Processing: The FPGA architecture allows for parallel "shift calculations," significantly speeding up the matrix multiplications inherent in neural networks.

Experimental Setup

The system was evaluated using the Cornell Grasping Dataset, a standard benchmark in the field.

  • Dataset Size: 1,802 total image patches (1,442 for training, 360 for testing).
  • Input: 24x24 pixel depth image patches (576 total features per patch).
  • Labels: Binary classification—either a "positive" (successful) or "negative" (unsuccessful) grasp rectangle.

Key Results and Insights

  • High Accuracy: The proposed SCN-AE model achieved an average recognition accuracy of 91.935% on the FPGA.
  • Superior Performance: SCN-AE outperformed both the standard SCN (~89.3%) and the traditional ELM (~81.2%), proving that the autoencoder layer is vital for handling depth images.
  • Hardware Fidelity: There was only a ~0.3% difference in accuracy between the MATLAB simulation and the physical FPGA implementation, confirming that the bit-width and function approximations were highly effective.
  • Stability: The SCN-AE showed much lower variance across different training runs compared to ELM, making it a more reliable choice for real-world robotics.

The "aha" moment for SCN is that randomness doesn't have to be blind. By supervising random weight generation with an inequality constraint, the network achieves the speed of a random network with the mathematical rigor of a universal approximator. Furthermore, the paper demonstrates that a "shallow" autoencoder can effectively compress depth data without the massive overhead of deep learning stacks.

Conclusion

While the results are promising, future work includes testing on larger, more diverse datasets (like Jacquard) to further prove generalization. The next step is integrating this FPGA framework into a full "closed-loop" robotic system where the vision system directly drives the motor controllers of a physical arm, particularly in dynamic environments with moving objects.

Summary updated

Abstract

Precise and rapid grasping recognition based on machine vision is one of the challenging problems for intelligent robots. As a nonlinear network for recognition, the stochastic configuration network (SCN) is considered as a promising method due to its universal approximation capability and fast modelling property, therefore, in this paper, SCN with depth image is applied to robotic grasping recognition for the first time. Moreover, in order to obtain better grasping performance, a novel autoencoder-based SCN (SCN-AE) is proposed to extract the higher level features of the captured image. Finally, a field programmable gate array (FPGA) framework for implementing SCN models is designed and analyzed, which can effectively balance the recognition accuracy and resource utilization. Experimental results on Cornell grasping dataset proved that the proposed SCN-AE method achieved higher accuracy and better stability in robotic grasping recognition task, and the designed FPGA framework implemented the SCN and SCN-AE recognition models successfully. The average recognition accuracy of the proposed SCN-AE based on FPGA reaches 91.935%.
BibTeX
@inproceedings{pan2020fpga,
  author = {Jiaqi Pan and Feng Luan and Yunqi Gao and Yangjie Wei},
  title = {{FPGA-Based Implementation of Stochastic Configuration Network for Robotic Grasping Recognition}},
  booktitle = {IEEE Access},
  year = {2020},
  url = {https://openreview.net/forum?id=MtR1e5cHHx},
}
Figure