No abstract

Occlusions, restricted field of view and limited resolution all constrain a robot's ability to sense its environment from a single observation. In these cases, the robot first needs to actively query multiple observations and accumulate information before it can complete a task. In this paper, we cast this problem of active vision as active inference, which states that an intelligent agent maintains a generative model of its environment and acts in order to minimize its surprise, or expected free energy according to this model. We apply this to an object-reaching task for a 7-DOF robotic manipulator with an in-hand camera to scan the workspace. A novel generative model using deep neural networks is proposed that is able to fuse multiple views into an abstract representation and is trained from data by minimizing variational free energy. We validate our approach experimentally for a reaching task in simulation in which a robotic agent starts without any knowledge about its workspace. Each step, the next view pose is chosen by evaluating the expected free energy. We find that by minimizing the expected free energy, exploratory behavior emerges when the target object to reach is not in view, and the end effector is moved to the correct reach position once the target is located. Similar to an owl scavenging for prey, the robot naturally prefers higher ground for exploring, approaching its target once located.

Scene understanding and decomposition is a crucial challenge for intelligent systems, whether it is for object manipulation, navigation, or any other task. Although current machine and deep learning approaches for object detection and classification obtain high accuracy, they typically do not leverage interaction with the world and are limited to a set of objects seen during training. Humans on the other hand learn to recognize and classify different objects by actively engaging with them on first encounter. Moreover, recent theories in neuroscience suggest that cortical columns in the neocortex play an important role in this process, by building predictive models about objects in their reference frame. In this article, we present an enactive embodied agent that implements such a generative model for object interaction. For each object category, our system instantiates a deep neural network, called Cortical Column Network (CCN), that represents the object in its own reference frame by learning a generative model that predicts the expected transform in pixel space, given an action. The model parameters are optimized through the active inference paradigm, i.e., the minimization of variational free energy. When provided with a visual observation, an ensemble of CCNs each vote on their belief of observing that specific object category, yielding a potential object classification. In case the likelihood on the selected category is too low, the object is detected as an unknown category, and the agent has the ability to instantiate a novel CCN for this category. We validate our system in an simulated environment, where it needs to learn to discern multiple objects from the YCB dataset. We show that classification accuracy improves as an embodied agent can gather more evidence, and that it is able to learn about novel, previously unseen objects. Finally, we show that an agent driven through active inference can choose their actions to reach a preferred observation.

Although modern object detection and classification models achieve high accuracy, these are typically constrained in advance on a fixed train set and are therefore not flexible to deal with novel, unseen object categories. Moreover, these models most often operate on a single frame, which may yield incorrect classifications in case of ambiguous viewpoints. In this paper, we propose an active inference agent that actively gathers evidence for object classifications, and can learn novel object categories over time. Drawing inspiration from the human brain, we build object-centric generative models composed of two information streams, a what-and a where-stream. The what-stream predicts whether the observed object belongs to a specific category, while the where-stream is responsible for representing the object in its internal 3D reference frame. We show that our agent (i) is able to learn representations for many object categories in an unsupervised way, (ii) achieves state-of-theart classification accuracies, actively resolving ambiguity when required and (iii) identifies novel object categories. Furthermore, we validate our system in an end-to-end fashion where the agent is able to search for an object at a given pose from a pixel-based rendering. We believe that this is a first step towards building modular, intelligent systems that can be used for a wide range of tasks involving three dimensional objects.

The robotics field has been deeply influenced by the advent of deep learning. In recent years, this trend has been characterized by the adoption of large, pretrained models for robotic use cases, which are not compatible with the computational hardware available in robotic systems. Moreover, such large, computationally intensive models impede the low-latency execution which is required for many closed-loop control systems. In this work, we propose different strategies for improving the computational efficiency of the deep-learning models adopted in reinforcement-learning (RL) scenarios. As a use-case project, we consider an image-based RL method on the synergy between push-and-grasp actions. As a first optimization step, we reduce the model architecture in complexity, by decreasing the number of layers and by altering the architecture structure. Second, we consider downscaling the input resolution to reduce the computational load. Finally, we perform weight quantization, where we compare post-training quantization and quantized-aware training. We benchmark the improvements introduced in each optimization by running a standard testing routine. We show that the optimization strategies introduced can improve the computational efficiency by around 300 times, while also slightly improving the functional performance of the system. In addition, we demonstrate closed-loop control behaviour on a real-world robot, while processing everything on a Jetson Xavier NX edge device.