Intrinsic motivation, inspired by the psychology of developmental learning in infants, stimulates exploration in agents without relying solely on sparse external rewards. Existing methods in reinforcement learning like Random Network Distillation (RND) face significant limitations, including (1) relying on raw visual inputs, leading to a lack of meaningful representations, (2) the inability to build a robust latent space, (3) poor target network initialization and (4) rapid degradation of intrinsic rewards. In this paper, we introduce Pre-trained Network Distillation (PreND), a novel approach to enhance intrinsic motivation in reinforcement learning (RL) by improving upon the widely used prediction-based method, RND. PreND addresses these challenges by incorporating pre-trained representation models into both the target and predictor networks, resulting in more meaningful and stable intrinsic rewards, while enhancing the representation learned by the model. We also tried simple but effective variants of the predictor network optimization by controlling the learning rate. Through experiments on the Atari domain, we demonstrate that PreND significantly outperforms RND, offering a more robust intrinsic motivation signal that leads to better exploration, improving overall performance and sample efficiency. This research highlights the importance of target and predictor networks representation in prediction-based intrinsic motivation, setting a new direction for improving RL agents' learning efficiency in sparse reward environments.
We tackle the challenge of rapidly adapting an agent's behavior to solve spatiotemporally continuous problems in novel settings. Animals exhibit extraordinary abilities to adapt to new contexts, a capacity unmatched by artificial systems. Instead of focusing on generalization through deep reinforcement learning, we propose viewing behavior as the physical manifestation of a search procedure, where robust problem-solving emerges from an exhaustive search across all possible behaviors. Surprisingly, this can be done efficiently using online modification of a cognitive graph that guides action, challenging the predominant view that exhaustive search in continuous spaces is impractical. We describe an algorithm that implicitly enumerates behaviors by regulating the tight feedback loop between execution of behaviors and mutation of the graph, and provide a neural implementation based on Hebbian learning and a novel high-dimensional harmonic representation inspired by entorhinal cortex. By framing behavior as search, we provide a mathematically simple and biologically plausible model for real-time behavioral adaptation, successfully solving a variety of continuous state-space navigation problems. This framework not only offers a flexible neural substrate for other applications but also presents a powerful paradigm for understanding adaptive behavior. Our results suggest potential advancements in developmental learning and unsupervised skill acquisition, paving the way for autonomous robots to master complex skills in data-sparse environments demanding flexibility.
We introduce a generic, compositional and interpretable class of generative world models that supports open-ended learning agents. This is a sparse class of Bayesian networks capable of approximating a broad range of stochastic processes, which provide agents with the ability to learn world models in a manner that may be both interpretable and computationally scalable. This approach integrating Bayesian structure learning and intrinsically motivated (model-based) planning enables agents to actively develop and refine their world models, which may lead to developmental learning and more robust, adaptive behavior.
The mainstream AIs approaches are the generative and deep learning approaches with large language models (LLMs) and the manually constructed symbolic approach. Both approaches have led to valuable AI systems and impressive feats. However, manually constructed AIs are brittle even in circumscribed domains. Generative AIs make strange mistakes and do not notice them. In both approaches the AIs cannot be instructed easily, fail to use common sense, and lack curiosity. They have abstract knowledge but lack social alignment. Developmental AIs have more potential. They start with innate competences, interact with their environment, and learn from their interactions. They interact and learn from people and establish perceptual, cognitive, and common grounding. Developmental AIs have demonstrated capabilities including multimodal perception, object recognition, and manipulation. Powerful computational models for hierarchical planning, abstraction discovery, curiosity, and language acquisition exist but need to be adapted to a developmental learning based approach. The promise is that developmental AIs will acquire self-developed and socially developed competences. They would address the shortcomings of current mainstream AI approaches, and ultimately lead to sophisticated forms of learning involving critical reading, provenance evaluation, and hypothesis testing. However, developmental AI projects have not yet fully reached the Speaking Gap corresponding to toddler development at about two years of age, before their speech is fluent. The AIs do not bridge the Reading Gap, to skillfully and skeptically learn from written and online information resources. This position paper lays out the prospects, gaps, and challenges for extending the practice of developmental AIs to create resilient, intelligent, and human-compatible AIs that learn what they need to know.
There is much to learn through synthesis of Developmental Biology, Cognitive Science and Computational Modeling. Our path forward involves a design for developmentally-inspired learning agents based on Braitenberg Vehicles. Continual developmental neurosimulation allows us to consider the role of developmental trajectories in bridging the related phenomena of nervous system morphogenesis, developmental learning, and plasticity. Being closely tied to continual learning, our approach is tightly integrated with developmental embodiment, and can be implemented using a type of agent called developmental Braitenberg Vehicles (dBVs). dBVs begin their lives as a set of undefined structures that transform into agent-based systems including a body, sensors, effectors, and nervous system. This phenotype is characterized in terms of developmental timing: with distinct morphogenetic, critical, and acquisition (developmental learning) periods. We further propose that network morphogenesis can be accomplished using a genetic algorithmic approach, while developmental learning can be implemented using a number of computational methodologies. This approach provides a framework for adaptive agent behavior that might result from a developmental approach: namely by exploiting critical periods or growth and acquisition, an explicitly embodied network architecture, and a distinction between the assembly of neuronal networks and active learning on these networks. In conclusion, we will consider agent learning and development at different timescales, from very short (<100ms) intervals to long-term evolution. The development, evolution, and learning in an embodied agent-based approach is key to an integrative view of biologically-inspired intelligence.
Connecting brain and behavior is a longstanding issue in the areas of behavioral science, artificial intelligence, and neurobiology. As is standard among models of artificial and biological neural networks, an analogue of the fully mature brain is presented as a blank slate. However, this does not consider the realities of biological development and developmental learning. Our purpose is to model the development of an artificial organism that exhibits complex behaviors. We introduce three alternate approaches to demonstrate how developmental embodied agents can be implemented. The resulting developmental BVs (dBVs) will generate behaviors ranging from stimulus responses to group behavior that resembles collective motion. We will situate this work in the domain of artificial brain networks along with broader themes such as embodied cognition, feedback, and emergence. Our perspective is exemplified by three software instantiations that demonstrate how a BV-genetic algorithm hybrid model, multisensory Hebbian learning model, and multi-agent approaches can be used to approach BV development. We introduce use cases such as optimized spatial cognition (vehicle-genetic algorithm hybrid model), hinges connecting behavioral and neural models (multisensory Hebbian learning model), and cumulative classification (multi-agent approaches). In conclusion, we consider future applications of the developmental neurosimulation approach.
Evolution and development operate at different timescales; generations for the one, a lifetime for the other. These two processes, the basis of much of life on earth, interact in many non-trivial ways, but their temporal hierarchy -- evolution overarching development -- is observed for most multicellular lifeforms. When designing robots however, this tenet lifts: it becomes -- however natural -- a design choice. We propose to inverse this temporal hierarchy and design a developmental process happening at the phylogenetic timescale. Over a classic evolutionary search aimed at finding good gaits for tentacle 2D robots, we add a developmental process over the robots' morphologies. Within a generation, the morphology of the robots does not change. But from one generation to the next, the morphology develops. Much like we become bigger, stronger, and heavier as we age, our robots are bigger, stronger and heavier with each passing generation. Our robots start with baby morphologies, and a few thousand generations later, end-up with adult ones. We show that this produces better and qualitatively different gaits than an evolutionary search with only adult robots, and that it prevents premature convergence by fostering exploration. In addition, we validate our method on voxel lattice 3D robots from the literature and compare it to a recent evolutionary developmental approach. Our method is conceptually simple, and can be effective on small or large populations of robots, and intrinsic to the robot and its morphology, not the task or environment. Furthermore, by recasting the evolutionary search as a learning process, these results can be viewed in the context of developmental learning robotics.
In this paper, we present a novel developmental reinforcement learning-based controller for a quadcopter with thrust vectoring capabilities. This multirotor UAV design has tilt-enabled rotors. It utilizes the rotor force magnitude and direction to achieve the desired state during flight. The control policy of this robot is learned using the policy transfer from the learned controller of the quadcopter (comparatively simple UAV design without thrust vectoring). This approach allows learning a control policy for systems with multiple inputs and multiple outputs. The performance of the learned policy is evaluated by physics-based simulations for the tasks of hovering and way-point navigation. The flight simulations utilize a flight controller based on reinforcement learning without any additional PID components. The results show faster learning with the presented approach as opposed to learning the control policy from scratch for this new UAV design created by modifications in a conventional quadcopter, i.e., the addition of more degrees of freedom (4-actuators in conventional quadcopter to 8-actuators in tilt-rotor quadcopter). We demonstrate the robustness of our learned policy by showing the recovery of the tilt-rotor platform in the simulation from various non-static initial conditions in order to reach a desired state. The developmental policy for the tilt-rotor UAV also showed superior fault tolerance when compared with the policy learned from the scratch. The results show the ability of the presented approach to bootstrap the learned behavior from a simpler system (lower-dimensional action-space) to a more complex robot (comparatively higher-dimensional action-space) and reach better performance faster.
Taking inspiration from developmental learning, we present a novel reinforcement learning architecture which hierarchically learns and represents self-generated skills in an end-to-end way. With this architecture, an agent focuses only on task-rewarded skills while keeping the learning process of skills bottom-up. This bottom-up approach allows to learn skills that 1- are transferable across tasks, 2- improves exploration when rewards are sparse. To do so, we combine a previously defined mutual information objective with a novel curriculum learning algorithm, creating an unlimited and explorable tree of skills. We test our agent on simple gridworld environments to understand and visualize how the agent distinguishes between its skills. Then we show that our approach can scale on more difficult MuJoCo environments in which our agent is able to build a representation of skills which improve over a baseline both transfer learning and exploration when rewards are sparse.
CNNs have made an undeniable impact on computer vision through the ability to learn high-capacity models with large annotated training sets. One of their remarkable properties is the ability to transfer knowledge from a large source dataset to a (typically smaller) target dataset. This is usually accomplished through fine-tuning a fixed-size network on new target data. Indeed, virtually every contemporary visual recognition system makes use of fine-tuning to transfer knowledge from ImageNet. In this work, we analyze what components and parameters change during fine-tuning, and discover that increasing model capacity allows for more natural model adaptation through fine-tuning. By making an analogy to developmental learning, we demonstrate that "growing" a CNN with additional units, either by widening existing layers or deepening the overall network, significantly outperforms classic fine-tuning approaches. But in order to properly grow a network, we show that newly-added units must be appropriately normalized to allow for a pace of learning that is consistent with existing units. We empirically validate our approach on several benchmark datasets, producing state-of-the-art results.
Emotional concepts play a huge role in our daily life since they take part into many cognitive processes: from the perception of the environment around us to different learning processes and natural communication. Social robots need to communicate with humans, which increased also the popularity of affective embodied models that adopt different emotional concepts in many everyday tasks. However, there is still a gap between the development of these solutions and the integration and development of a complex emotion appraisal system, which is much necessary for true social robots. In this paper, we propose a deep neural model which is designed in the light of different aspects of developmental learning of emotional concepts to provide an integrated solution for internal and external emotion appraisal. We evaluate the performance of the proposed model with different challenging corpora and compare it with state-of-the-art models for external emotion appraisal. To extend the evaluation of the proposed model, we designed and collected a novel dataset based on a Human-Robot Interaction (HRI) scenario. We deployed the model in an iCub robot and evaluated the capability of the robot to learn and describe the affective behavior of different persons based on observation. The performed experiments demonstrate that the proposed model is competitive with the state of the art in describing emotion behavior in general. In addition, it is able to generate internal emotional concepts that evolve through time: it continuously forms and updates the formed emotional concepts, which is a step towards creating an emotional appraisal model grounded in the robot experiences.
Intrinsically motivated spontaneous exploration is a key enabler of autonomous developmental learning in human children. It enables the discovery of skill repertoires through autotelic learning, i.e. the self-generation, self-selection, self-ordering and self-experimentation of learning goals. We present an algorithmic approach called Intrinsically Motivated Goal Exploration Processes (IMGEP) to enable similar properties of autonomous learning in machines. The IMGEP architecture relies on several principles: 1) self-generation of goals, generalized as parameterized fitness functions; 2) selection of goals based on intrinsic rewards; 3) exploration with incremental goal-parameterized policy search and exploitation with a batch learning algorithm; 4) systematic reuse of information acquired when targeting a goal for improving towards other goals. We present a particularly efficient form of IMGEP, called AMB, that uses a population-based policy and an object-centered spatio-temporal modularity. We provide several implementations of this architecture and demonstrate their ability to automatically generate a learning curriculum within several experimental setups. One of these experiments includes a real humanoid robot exploring multiple spaces of goals with several hundred continuous dimensions and with distractors. While no particular target goal is provided to these autotelic agents, this curriculum allows the discovery of diverse skills that act as stepping stones for learning more complex skills, e.g. nested tool use.