Learning Probabilistic Multi-Modal Actor Models for Vision-Based Robotic Grasping
This addresses the computational bottleneck in robotic grasping for applications requiring real-time action, though it is incremental as it builds on existing value network approaches.
The paper tackles the problem of slow inference time in vision-based robotic grasping by proposing a neural density model to directly approximate successful grasp poses from images, achieving similar performance while reducing inference time by 3 times compared to the state-of-the-art method.
Many previous works approach vision-based robotic grasping by training a value network that evaluates grasp proposals. These approaches require an optimization process at run-time to infer the best action from the value network. As a result, the inference time grows exponentially as the dimension of action space increases. We propose an alternative method, by directly training a neural density model to approximate the conditional distribution of successful grasp poses from the input images. We construct a neural network that combines Gaussian mixture and normalizing flows, which is able to represent multi-modal, complex probability distributions. We demonstrate on both simulation and real robot that the proposed actor model achieves similar performance compared to the value network using the Cross-Entropy Method (CEM) for inference, on top-down grasping with a 4 dimensional action space. Our actor model reduces the inference time by 3 times compared to the state-of-the-art CEM method. We believe that actor models will play an important role when scaling up these approaches to higher dimensional action spaces.