Ing. Miloš Prágr

In this study, we address generalized autonomous mobile robot exploration of unknown environments where a robotic agent learns a traversability model and builds a spatial model of the environment. The agent can benefit from the model learned online in distinguishing what terrains are easy to traverse and which should be avoided. The proposed solution enables the learning of multiple traversability models, each associated with a particular locomotion gait, a walking pattern of a multi-legged walking robot. We propose to address the simultaneous learning of the environment and traversability models by a decoupled approach. Thus, navigation waypoints are generated using the current spatial and traversability models to gain the information necessary to improve the particular model during the robot's motion in the environment. From the set of possible waypoints, the decision on where to navigate next is made based on the solution of the generalized traveling salesman problem that allows taking into account a planning horizon longer than a single myopic decision. The proposed approach has been verified in simulated scenarios and experimental deployments with a real hexapod walking robot with two locomotion gaits, suitable for different terrains. Based on the achieved results, the proposed method exploits the online learned traversability models and further supports the selection of the most appropriate locomotion gait for the particular terrain types.

For autonomous robots operating in an unknown environment, it is important to assess the traversability of the surrounding terrain to improve path planning and decision-making on where to navigate next in a cost-efficient way. Specifically, in mobile robot exploration, terrains and their traversability are unknown prior to the deployment. The robot needs to use its limited resources to learn its terrain traversability model on the go; however, reusing a provided model is still a desirable option. In a team of heterogeneous robots, the models assessing traversability cannot be reused directly since robots might possess different morphology or sensory equipment and thus experience the terrain differently. In this paper, we propose a transfer learning approach for convolutional neural networks assessing the traversability between heterogeneous robots, where the transferred network is retrained using data available for the target robot to accommodate itself to the robot’s traversability. The proposed method is verified in real-world experiments, where the proposed approach provides faster learning convergence and better traversal cost predictions than the baseline.

The terrain traversal abilities of multi-legged walking robots are affected by gaits, the walking patterns that enable adaptation to various operational environments. Fast and lowset gaits are suited to flat ground, while cautious and highset gaits enable traversing rough areas. A suitable gait can be selected using prior experience with a particular terrain type. However, experience alone is insufficient in practical setups, where the robot experiences each terrain with only one or just a few gaits and thus would infer novel gait-terrain interactions from insufficient data. Therefore, we use knowledge transfer to address unsampled gait-terrain interactions and infer the traversal cost for every gait. The proposed solution combines gaitterrain cost models using inferred gait-to-gait models projecting the robot experiences between different gaits. We implement the cost models as Gaussian Mixture regressors providing certainty to identify unknown terrains where knowledge transfer is desirable. The presented method has been verified in synthetic showcase scenarios and deployment with a real walking robot. The proposed knowledge transfer demonstrates improved cost prediction and selection of the appropriate gait for specific terrains.

Predicting mobile robots' traversability over terrains is crucial to select safe and efficient paths through rough and unstructured environments. In multi-robot missions, knowledge transfer techniques can enable learning terrain traversability assessment the robots did not experience individually. The knowledge can be incrementally aggregated for homogeneous robots since they can treat foreign knowledge as their own. However, robots with different perceptions might experience the same terrain differently, so it is impossible to aggregate the shared knowledge directly. In this paper, we show how to learn a model that transfers the experience between heterogeneous robots, enabling each robot to use the whole sum of the experience of the multi-robot team. The proposed approach uses correlation to combine individual neural networks that assess the traversability of individual robots. The presented method has been verified in a real-world deployment of multi-legged walking robots with different cost assessment policies.

Communication is of crucial importance for coordinating a team of mobile robotic units. In environments such as underground tunnels, the propagation of wireless signals is affected by nontrivial physical phenomena. Hence, both modeling of the communication properties and the consequent task to estimate where communication is available becomes demanding. A communication map is a tool assessing the characteristic of communication between two arbitrary spatial coordinates. The existing approaches based on interpolation of a priori obtained spatial measurements do not provide precise extrapolation estimates for unvisited locations. Therefore, we propose to address the extrapolation of the signal strength by a position-independent model based on approximating the obstacle occupancy ratio between the signal source and receiver. The proposed approach is compared to the existing attenuation models based on free-space path loss and spatial projection using a natural cave dataset. Based on the reported results, the proposed approach provides more accurate predictions than the existing approaches.

A repetitive movement pattern of many animals, a gait, is controlled by the Central Pattern Generator (CPG), providing rhythmic control synchronous to the sensed environment. As a rhythmic signal generator, the CPG can control the motion phase of biomimetic legged robots without feedback. The CPG can also act in sensory synchronization, where it can be utilized as a sensory phase estimator. Direct use of the CPG as the estimator is not common, and there is little research done on its utilization in the phase estimation. Generally, the sensory estimation augments the sensory feedback information, and motion irregularities can reveal from comparing measurements with the estimation. In this work, we study the CPG in the context of phase irregularity detection, where the timing of sensory events is disturbed. We propose a novel self-supervised method for learning mistiming detection, where the neural detector is trained by dynamic Hebbian-like rules during the robot walking. The proposed detector is composed of three neural components: (i) the CPG providing phase estimation, (ii) Radial Basis Function neuron anticipating the sensory event, and (iii) Leaky Integrate-and-Fire neuron detecting the sensory mistiming. The detector is integrated with the CPG-based gait controller. The mistiming detection triggers two reflexes: the elevator reflex, which avoids an obstacle, and the search reflex, which grasps a missing foothold. The proposed controller is deployed and trained on a hexapod walking robot to demonstrate the mistiming detection in real locomotion. The trained system has been examined in the controlled laboratory experiment and real field deployment in the Bull Rock cave system, where the robot utilized mistiming detection to negotiate the unstructured and slippery subterranean environment.

In this paper, we report on the developed system for assessment of ground unit terrain traversal cost using aerial reconnaissance of the expected mission environment. The system combines an aerial vehicle with ground robot terrain learning in the traversal cost modeling utilized in the mission planning for ground units. The aerial vehicle is deployed to capture visual data used to build a terrain model that is then used for the extraction of the terrain features of the expected operational area of the ground units. Based on the previous traversal experience of the ground units in similar environments, the learned model of the traversal cost is employed to predict the traversal cost of the new expected operational area to plan a cost-efficient path to visit the desired locations of interest. The particular modules of the system are demonstrated in an experimental scenario combining the deployment of an unmanned aerial vehicle with a multi-legged walking robot used for learning the traversal cost model.

In this paper, we concern learning of terrain types based on the traversal experience observed by a hexapod walking robot. The addressed problem is motivated by the navigation of unmanned ground vehicles in long-term autonomous missions in a priory unknown environments such as extraterrestrial exploration. In such deployments, the robotic vehicle needs to learn hard to traverse terrains to improve its autonomous performance and avoid possibly dangerous areas. We propose to utilize Growing Neural Gas for terrain learning to capture the robot experience with traversing the terrain and thus learn a classifier of individual terrain types. The classifier is learned using a real time-series dataset collected by a hexapod walking robot traversing various terrain types. The learned model can be utilized to predict the traversal cost of newly observed terrains to support decisions on where to navigate next.

In multi-robot deployments, the robots need to share and integrate their own experience and perform transfer learning. Under the assumption that the robots have the same morphology and carry equivalent sensory equipment, the problem of transfer learning can be considered incremental learning. Thus, the transfer learning problem inherits the challenges of incremental learning, such as catastrophic forgetting and concept drift. In catastrophic forgetting, the model abruptly forgets the previously learned knowledge during the learning process. The concept drift arises with different experiences between consecutively sampled models. However, state-of-the-art robotic transfer learning approaches do not address both challenges at once. In this paper, we propose to use an incremental classifier on a transfer learning problem. The feasibility of the proposed approach is demonstrated in a real deployment. The robot consistently merges two classifiers learned on two different tasks into a classifier that performs well on both tasks.

Motivated by the deployment of multi-legged walking robots in traversing various terrain types, we benchmark existing online and unsupervised incremental learning approaches in traversal cost prediction. The traversal cost is defined by the proprioceptive signal of the robot traversal stability that is combined with appearance and geometric properties of the traversed terrains to construct the traversal cost model incrementally. In the motivational deployment, such a model is instantaneously utilized to extrapolate the traversal cost for observed areas that have not yet been visited by the robot to avoid difficult terrains in motion planning. The examined approaches are Incremental Gaussian Mixture Network, Growing Neural Gas, Improved Self-Organizing Incremental Neural Network, Locally Weighted Projection Regression, and Bayesian Committee Machine with Gaussian Process Regressors. The performance is examined using a dataset of the various terrains traversed by a real hexapod walking robot. A part of the presented benchmarking is thus a description of the dataset and also a construction of the reference traversal cost model that is used for comparison of the evaluated regressors. The reference is designed as a compound Gaussian process-based model that is learned separately over the individual terrain types. Based on the evaluation results, the best performance among the examined regressors is provided by Incremental Gaussian Mixture Network, Improved Self-Organizing Incremental Neural Network, and Locally Weighted Projection Regression, while the latter two have the lower computational requirements.

In this paper, we address traversability cost estimation using exteroceptive and proprioceptive data collected by a team of aerial and ground vehicles. The main idea of the proposed approach is to estimate the terrain traversability cost based on the real experience of the multi-legged walking robot with traversing different terrain types. We propose to combine visual features with the real measured traversability cost based on proprioceptive signals of the utilized hexapod walking robot as a ground unit. The estimated traversability cost is augmented by extracted visual features from the onboard robot camera, and the features are utilized to extrapolate the learned traversability model for an aerial scan of new environments to assess their traversability cost. The extrapolated traversability cost can be utilized in the high-level mission planning to avoid areas that are difficult to traverse but not visited by the ground units. The proposed approach has been experimentally verified with a real hexapod walking robot in indoor and outdoor scenarios.

In this paper, we report early results on the deployment of the growing neural gas algorithm in online incremental learning of traversability assessment with a multi-legged walking robot. The addressed problem is to incrementally build a model of the robot experience with traversing the terrain that can be immediately utilized in the traversability cost assessment of seen but not yet visited areas. The main motivation of the studied deployment is to improve the performance of the autonomous mission by avoiding hard to traverse areas and support planning cost-efficient paths based on the continuously collected measurements characterizing the operational environment. We propose to employ the growing neural gas algorithm to incrementally build a model of the terrain characterization from exteroceptive features that are associated with the proprioceptive based estimation of the traversal cost. Based on the reported results, the proposed deployment provides competitive results to the existing approach based on the Incremental Gaussian Mixture Network.

This paper reports on the deployment of self-organizing maps in unsupervised learning of the traversal cost for a hexapod walking robot. The problem is motivated by traversability assessment of terrains not yet visited by the robot, but for which shape and appearance features are available. The perception system of the robot is used to extract terrain features that are accompanied by traversal cost characterization captured from the real experience of the robot with the terrain, which is characterized by proprioceptive features. The learned model is employed to predict the traversal cost of new terrains based only on the shape and appearance features. Based on the experimental deployment of the robot in various terrains, a dataset of the traversal cost has been collected that is utilized in the presented evaluation of the traversal cost modeling using self-organizing map approach. In comparison with the Gaussian process, the self-organizing map provides competitive results and the found paths using the predicted traversal costs are close to the optimal path based on reference traversal cost of the particular terrain types. Besides, the self-organizing map can also be utilized for unsupervised identification of the terrain types, and it further supports incremental learning, which is more suitable for practical deployments of the robot in a priory unknown environments where reference traversal costs are not available.

In this paper, we address motion efficiency in autonomous robot exploration with multi-legged walking robots that can traverse rough terrains at the cost of lower efficiency and greater body vibration. We propose a robotic system for online and incremental learning of the terrain traversal cost that is immediately utilized to reason about next navigational goals in building spatial model of the robot surrounding. The traversal cost experienced by the robot is characterized by incrementally constructed Gaussian Processes using Bayesian Committee Machine. During the exploration, the robot builds the spatial terrain model, marks untraversable areas, and leverages the Gaussian Process predictive variance to decide whether to improve the spatial model or decrease the uncertainty of the terrain traversal cost. The feasibility of the proposed approach has been experimentally verified in a fully autonomous deployment with the hexapod walking robot

In this paper, we concern a traversal cost estimation considering motion control of a hexapod walking robot. The proposed idea is motivated by the observation that the traversal cost depends not only on the traversed terrain but also on the robot motion. Based on the experimental deployments, the forward motion is preferable over some terrains; however, uphill and downhill locomotion over the particular terrain might differ significantly. Therefore, we propose to enhance the traversal cost model by a motion characterization. The model is learned using feature descriptor composed of terrain shape and appearance that is combined with the expected motion performance determined from the slope change and possible rotation of the robot. The traversal model enables to reason about the robot stability regarding placement of the robot legs and performed motion action. The proposed idea of motion characterization is demonstrated and experimentally verified on a simplified motion control using grid-based planning with the robot control decomposed into straight and turn movements.

The effectiveness of the robot locomotion can be measured using the cost of transport (CoT) which represents the amount of energy that is needed for traversing from one place to another. Terrains excerpt different mechanical properties when crawled by a multi-legged robot, and thus different values of the CoT. It is therefore desirable to estimate the CoT in advance and plan the robot motion accordingly. However, the CoT might not be known prior the robot deployment, e.g., in extraterrestrial missions; hence, a robot has to learn different terrains as it crawls through the environment incrementally. In this work, we focus on estimating the CoT from visual and geometrical data of the crawled terrain. A thorough analysis of different terrain descriptors within the context of incremental learning is presented to select the best performing approach. We report on the achieved results and experimental verification of the selected approaches with a real hexapod robot crawling over six different terrains.