In this paper, we present a novel robot model of touch, and its representation in an artificial cortex, that aims to capture some of the complexity of human touch. In particular, our approach integrates artificial mechanoception and nociception in an adaptive sensory field (the robot's "sensory body"), allowing for a more comprehensive simulation of tactile sensations. The robot's sensory field is then processed by a biologically plausible neural network in a way akin to sensory processing in the somatosensory and anterior cingulate cortex. Findings from our experimental results show our model's ability to integrate complex data from infrared sensors, leading to the emergence of a spatial sensory body representation in our neural network, with potentially significant implications for robot perception and interaction.
Observations in biologically inspired swarm formations from nature, like flocks of birds, herds of mammals, and packs of wolves, have inspired the innovation of various multi-robotic architectures. This work presents a robotic system that mimics leader-follower behaviors in the navigation and formation of sparse and dense environments. This work extends the original work by Weitzenfeld et al. to evaluate new swarm-based multi-robot architectures with obstacle avoidance and variations in group formations. The multiple robot architecture is based on a wolf pack with a defined 'alpha wolf,' which acts as the leader, and defines'betas wolves,' which act as followers. The 'alpha wolf' leads multiple 'beta wolves' that follow in formation behind the lead wolf, keeping track of a group member and maintaining a set angle and distance while performing obstacle avoidance, staying in formation, and performing speed adjustment. Variations in swarm formation behaviors being analyzed with robots include (1) beta robots following the alpha robot, (2) beta robots following the closest neighboring robot, and (3) robots following the same robot identified since the beginning. Experiments are performed in simulation, using Webots, to analyze robot formations.
The collective escape of predators by prey is a classic example of adaptive behavior in animal groups. Across species, prey has evolved a large repertoire of individual evasive maneuvers they can use to evade predators. With recent technological advances, more empirical data of collective escape is becoming available, and a large variation in the collective dynamics of different species is apparent. However, given the complexity of patterns of collective escape, we are still lacking the tools to understand their emergence. Computational models that can link rules of individual behavior to patterns of collective escape are needed, but species-specific motion and escape characteristics that will allow the link between behavior and eco-evolutionary dynamics of a given species are rarely included in agent-based models of collective behavior. Here, to tackle this challenge, we introduce a framework that uses individual-based state machines to model spatio-temporal dynamics of collective escape. A synthetic agent in our framework can switch its behavior between 'flocking' with different coordination specifics (e.g., quicker interactions when vigilant) and 'escape' with various maneuvers through a dynamic Markov-chain, depending on its local information (e.g., its relative position to the predator). A user can compose a new agent-based model adjusted to empirical data by choosing a set of states (which includes rules of motion, interaction, and escape), their temporal order, and a detailed parameterization. The flexibility and structure of our software allows substantial changes in a model with very minimal code alterations, showing great potential for future use to identify the underlying mechanisms of collective escape across species and ecological contexts.
In collective biological systems, social contagion processes play an important role in evaluating and processing information on the level of the collective. These group-level abilities typically arise from individual-level mechanisms and through local interactions. We are interested in the role of these mechanisms and their effect on the system's response to environmental inputs. In this paper, we present a spatially embedded network model that is inspired by large fish shoals performing collective action in response to predation. We compare the observations of spatio-temporal dynamics in the model simulations with empirical observations, studying specifically the effect of spatial heterogeneities on system activity. The model demonstrates how already simple mechanisms suffice to represent key characteristics of the study system, and highlights the importance of taking into account the spatial embedding for understanding group-level processes in animal collectives.
With the rise of robot swarms, it has become a relevant problem how humans can control them. Extended swarming is a potential approach in which robot swarms are treated as self-organising extensions of human bodies. Swarm control takes the form of controlling the observable swarm body while robot chains connect the human operator to relevant aspects of the environment. Inspired by how natural bodies are controlled by a nervous system, we here investigate how the swarm body's self-organisation can be influenced by robot chains acting as embodied neural traces while remaining under human high-level control. Three design principles are proposed for such embodied neural computation. First, the swarm body's self-organisation is controlled both by top-down human control and bottom-up sensor inputs alike to the hierarchical control architecture of the nervous system. Second, robots participating in robot chains are treated as rate-coded neurons rendering the chains as embodied neural traces which offers intuitive control possibilities for the human. Third, neural and swarm self-organisation are integrated by utilizing the swarm's communication network as a scaffolding for neural function influencing swarm dynamics. This process is interpreted as embodied Hebbian learning. Human control over the swarm is demonstrated in a grid-based search-and-rescue simulation with the objective of selecting the most valuable subregion defined by accumulated victims in need. We evaluate how using embodied trace relevance in terms of neural activation improves completion time to finding the highest-value trace as well as how attracting units to relevant traces increases their robustness.
The use of humanoid robots within the field of neuroscience has gained substantial interest in recent years, specifically as a means to implement and assess biological concepts. This conceptual paper addresses the vital challenge of uncertainty in the context of lifelong bio-inspired sensorimotor learning. Inspired by insights from developmental and neuroscientific studies, we examine the role of self-learning, exploration, and coordination dynamics. Building on principles derived from neural mechanisms and the concept of brain plasticity, we represent a robot's internal sensorimotor model and its synaptic-like reorganizational changes through dynamic self-organizing maps. We propose a concept that builds on that and distinguishes itself by employing visuo-arm coordination not as an end goal with potential emergent behaviors, but as a feedback controller, emphasizing the integration of an explainable memory-embedded model for continuous sensorimotor self-learning. We illustrate the framework's potential in dynamic scenarios such as tool use, where enhanced adaptability and fast task resumption after motor perturbations or recurring tool changes provide significant benefits. Verifying a memory entry is significantly quicker than updating the visuo-motor model. Through the concept of a memory-embedded controller, we establish the groundwork for effective and lifelong learning of sensorimotor skills in humanoid robots.
Visual navigation through complex environments is a challenging task, yet ants navigate in them easily and accurately with low-resolution vision and limited neural resources. Inspired by ants, we have developed a series of visual familiarity-based navigation algorithms for teach-and-repeat style navigation. These algorithms learn the egocentric visual appearance of the world on a training route and, during subsequent navigation, move in the direction that leads to the best match of the current view with one of the scenes encountered during training. Because they do not depend on accurate feature extraction or map building these algorithms use relatively unprocessed low-resolution panoramic views, making them computationally efficient. However, the computational cost of comparing the current view with all training images is still quite large and the algorithm can get confused if the path crosses itself. Here we develop and test novel algorithms where the agent uses sequence information to adaptively select a window of route memories to navigate with. This algorithm is shown to successfully navigate real-world routes through ant-like habitats, including a figure of 8-route, as well as a long route along a corridor, with all computation performed onboard the robot.
The CuttleBot project aspires to encapsulate the sophisticated behavior of cuttlefish in a neurorobot. The long-term goal is to construct a machine that mirrors the unique intelligent behavior demonstrated by this invertebrate. The current CuttleBot prototype represents an early step towards realizing a robotic system capable of advanced environmental interaction and decision-making. Its custom-made shell demonstrates the camouflaging and signaling observed in cephalopods in response to environmental stimuli. Similar to cuttlefish, the CuttleBot hunts for prey and responds to predators with defensive behaviors. Cuttlefish are impressive learners. Therefore, reinforcement learning was implemented to learn the appropriate behavioral responses to predators (e.g., camouflage or hide) and prey (e.g., confuse and attack). By creating cognitive systems with insights from the natural world, the CuttleBot project lays the groundwork for an era of robotics that comprehends and interacts with the environment in ways that are as dynamic and complex as the biological entities that inspire it.
Habitat fragmentation is currently speeding up due to the impact of human activities. This leads to higher variations in habitat qualities for different species, influencing their behaviours. One key behaviour directly linked to species survival is dispersal (displacement from one habitat patch to another). In order to take successful dispersal decisions, animals rely on different sources of information. Private information is derived from the physical environment and is directly linked to its quality, whereas social information is derived from the behaviour of conspecifics. Few modelling studies include both information types and their associated acquisition costs in their modelling frameworks. We fill this gap by adding genetic factors influencing the evolution of information acquisition to an existing agent-based model. By varying total and relative acquisition costs for different environmental conditions and perceptual ranges, we show that dispersal strategies and information usage are heavily influenced by the type of environment and information acquisition costs. As the total cost of information rises, the use of information progressively disappears under all environmental conditions. In stable environments with a low cost of information, the acquisition of the cheapest type of information results in an increase in fitness. In environments where patch quality varies greatly, the type of information used also depends on the perceptual range of the agents: agents with a restricted perceptual range often select both types of information while, agents with a larger perceptual range almost exclusively use private information.
Human exploration, a cornerstone of our ability to solve novel problems, is a complex process, posing significant research challenges. Most previous studies simplify tasks to isolate specific variables, creating artificial problems that do not align with those humans have evolved to solve, thus limiting the generalizability of findings. To address this gap, we introduce the Lockbox paradigm: a novel, ecologically valid, and challenging task that promotes active exploration and physical interaction. Data from 91 participants interacting with the Lockbox reveal a remarkable human ability to adapt and solve problems efficiently in complex scenarios. By comparing different interaction methods, we demonstrate the critical role of cost variations, such as physical and cognitive costs, in driving attentiveness and shaping exploration strategies. These findings provide valuable insights into human exploration strategies, with potential applications in fields such as robotics and artificial intelligence.
We introduce a navigation algorithm inspired by directional sensitivity observed in CA1 place cells of the rat hippocampus. These cells exhibit directional polarization characterized by vector fields converging to specific locations in the environment, known as ConSinks [8]. By sampling from a population of such cells at varying orientations, an optimal vector of travel towards a goal can be determined. Our proposed algorithm aims to emulate this mechanism for learning goal-directed navigation tasks. We employ a novel learning rule that integrates environmental reward signals with an eligibility trace to determine the update eligibility of a cell's directional sensitivity. Compared to state-of-the-art Reinforcement Learning algorithms, our approach demonstrates superior performance and speed in learning to navigate towards goals in obstacle-filled environments. Additionally, we observe analogous behavior in our algorithm to experimental evidence, where the mean ConSink location dynamically shifts toward a new goal shortly after it is introduced.
Integrated information, denoted as , quantifies the intrinsic information within causal systems. Despite its profound theoretical implications, applications of have mostly taken place in simulations of arbitrary systems, particularly in terms of biological realism. This study applies calculations to biologically inspired robotic agents that adapt to environmental conditions, thus providing a novel context for observing changes in information integration. The agents’ neural network is evolved to demonstrate behavior similar to Braitenberg’s Vehicles. The neuro-mechanical design of these evolved agents are then suitable for analysis. Interestingly, early generations had higher values. In later generations the diversity of connection weights and the values decreased, leading to simpler and more reactive neural activations.
Prediction is an important foundation of cognitive and intelligent behavior. Recent advances in deep learning heavily depend on prediction, in the form of self-supervised learning based on prediction and reinforcement learning (reward prediction). However, how such predictive capabilities emerged from simple organisms has not been investigated fully. Prior works have shown the relationship between input delay and predictive function to compensate for such delay. In this paper, we investigate other key factors that may contribute to the emergence of predictive behavior in evolving neural networks. We set up a delayed reaching task with a two-segment articulated arm. The arm is controlled to reach a moving target, where the target’s coordinate information is received with a delay. Following our previous work, we introduced a tool to extend the reach, when the target is beyond the arm’s reach. In this task, without predicting the trajectory of the moving target, the controller cannot reach the target. For the controller, we used the NeuroEvolution of Augmenting Topologies (NEAT) algorithm. Our results indicate that an important (auxiliary) fitness criterion for the emergence of predictive behavior is that of reduced energy usage (in the form of economy of motion). Further analysis shows that the number of recurrent loops correlates with target reaching performance, but more strongly so with the energy constraint. We expect our findings to lead to further investigations on the role of energy constraints on the evolution of predictive behavior.
We propose a simulation-based model of flower finding in echolocating nectarivorous bats. In particular, we propose a behavior-based model that uses two sensorimotor loops to dock with flowers. The EchoVr, as we have termed our echo simulator, uses a bank of echoes collected by ensonifying real objects with a physical (bat-like) sonar device. Using the EchoVr, we built a 2D environment consisting of simulated objects. We trained a neural network to activate the correct sensorimotor loop based on the echoes received by the simulated bat. The model guides the simulated bat to dock successfully with the flower opening (95% success rate) by computing control commands solely from echoic inputs.
Inspired by recent human studies, this paper investigates the benefits of employing varying navigation strategies in robot teams. We explore how mixed navigation strategies impact task completion time, environment exploration, and overall system effectiveness in multi-robot systems. Experiments were conducted in a simulated rectangular environment using Clearpath PR2 robots and evaluated different navigation strategies observed in humans: 1) Route (RT) knowledge where agents follow a predefined path, 2) Survey (SW) knowledge where agents take the shortest path while avoiding obstacles, 3) Mixed strategies with varying proportions, such as 40
The coordinated circular motion of individuals within a group, known as milling, is a widely observed collective motion pattern across biological and artificial systems. However, existing models focused on achieving stable, albeit unnatural, patterns, while overlooking the embodiment aspect of real-world agents. Here, we employ a spatially explicit agent-based model with visual occlusions and a collision avoidance mechanism to address this gap and investigate the emergence of temporary milling states over time. We show that short yet frequent milling dynamics are prevalent in a distinct parameter region, characterised by a qualitative shift in group behaviour between dynamical regimes, suggesting adaptability. We also show that such milling states require a minimal field of vision that not only promotes their occurrence but also matches the typical field of view observed in biological systems.
Efficient outdoor navigation remains a challenge for autonomous robots, yet bees excel in robust long-range navigation with minimal computational resources. To do so, they scaffold learning through innate behaviours such as survey flights: loops centred on the nest to explore the environment, which they perform before foraging. While the 2D positions of these flights have been tracked by radar, it has not been tested how well these flights can support subsequent long-range visual homing, nor whether the 3D structure (not captured by the radar) has an effect on homing performance. Using a 6 km^2 3D LIDAR scan of the Rothamsted Research Center – where bumblebee flights were tracked in radar experiments – we recreate the trajectory of bumblebee exploration and foraging flights. We then render panoramic views of the visited coordinates, and use these to test the efficacy of visual homing over large distances, and flying altitudes ranging from 2 to 32 m above the ground. We find that our model can predict the direction of the target from up to 300 m. Additionally, homing improves at higher altitudes, but there is limited transferability of information between flying heights.
This study employs a modified formulation of the secon-dorder Drift-Diffusion model to investigate how the interaction between environmental and social cues influences individual decision-making in a binary choice scenario. Environmental information is represented as stochastic cues, often biased towards one of the choices, while social information is conveyed through signals from a group of identical agents making random decisions. The model incorporates simplified human perceptual characteristics via a visual network of social interactions, which considers perceptual limitations due to physical distances and visual occlusions. Model parameters and assumptions are informed by an ongoing behavioural experiment on behavioural contagion, conducted in human and artificial multi-agent systems using virtual reality. The stochastic evolution of decision states in response to environmental and social input mirrors the behavioural choices of human participants, who respond to stimuli presented in the virtual reality environment and social cues from a group of virtual agents. Manipulating the size and density of the group revealed that larger group sizes and lower densities lead to greater alignment of individual decisions with social cues, accompanied by shorter and more homogeneous response times and reduced accuracy. These findings afford preliminary insights into the behavioural experiment. With reciprocal informative exchange from experimental findings, this study would contribute to enhanced realism in future steps.
When a predator chases its prey, a mind game ensues, requiring both predator and prey to predict what the other will do next. These elements of uncertainty and opponency are also seen in analyses of realworld tasks and games. For instance, one way to define an optimal solution of a non-cooperative game is to find the Nash equilibrium, a state in which each agent in a game has optimized its strategy given the strategies of others. The Regularized Nash Dynamics (R-NaD) algorithm guarantees that policies will converge to the Nash equilibrium, creating AIs that beat top human players in tasks with hidden information. Our research compares the performance of deep reinforcement learning agents trained with and without R-NaD in a simple hide-and-seek game, aiming to see how well the agents process unknowns in the environment. We then apply explainable AI (XAI) techniques to the trained model to examine the kinds of information that trained policies encode about opponent strategies. We find that policies trained with R-NaD outperform policies trained in regular self-play when there is hidden information. Furthermore, R-NaD policies use their opponent's past positions to decide which actions to take, more so than regular self-play. These findings yield insights on how animals and artificial agents operate under spatial uncertainty.
Collective shepherding is a complex problem with potentially a broad range of applications. Its complexity arises from the interaction of two collectives: 'sheep' and 'shepherds', with the latter attempting to control and guide the 'sheep'. Here, we combine an agent-based model for the 'sheep'-flock with a heuristic algorithm for the adaptive behavior of shepherds with two different behavioral modes: collecting, i.e. keeping the sheep flock together, and driving the sheep towards the target. We show that this algorithm can achieve selforganized coordination among multiple shepherds without direct communication, and investigate how the shepherding performance depends on selected parameters of the system such as sheep flock size, number of shepherds, or parameters governing the switching between the shepherd behavioral modes. We demonstrate that the algorithm can also be applied to more challenging scenarios like controlling non-cohesive or passive agents without self-propulsion. Besides extending our understanding of collective shepherding, our model provides a starting point for future research into unexplored aspects of this complex dynamical behavior.
