Ville Kyrki

David Blanco-Mulero, Oriol Barbany, Gokhan Alcan et al.

Realistic physics engines play a crucial role for learning to manipulate deformable objects such as garments in simulation. By doing so, researchers can circumvent challenges such as sensing the deformation of the object in the realworld. In spite of the extensive use of simulations for this task, few works have evaluated the reality gap between deformable object simulators and real-world data. We present a benchmark dataset to evaluate the sim-to-real gap in cloth manipulation. The dataset is collected by performing a dynamic as well as a quasi-static cloth manipulation task involving contact with a rigid table. We use the dataset to evaluate the reality gap, computational time, and simulation stability of four popular deformable object simulators: MuJoCo, Bullet, Flex, and SOFA. Additionally, we discuss the benefits and drawbacks of each simulator. The benchmark dataset is open-source. Supplementary material, videos, and code, can be found at https://sites.google.com/view/cloth-sim2real-benchmark.

Ali Ghadirzadeh, Petra Poklukar, Karol Arndt et al. · stanford

We present a data-efficient framework for solving sequential decision-making problems which exploits the combination of reinforcement learning (RL) and latent variable generative models. The framework, called GenRL, trains deep policies by introducing an action latent variable such that the feed-forward policy search can be divided into two parts: (i) training a sub-policy that outputs a distribution over the action latent variable given a state of the system, and (ii) unsupervised training of a generative model that outputs a sequence of motor actions conditioned on the latent action variable. GenRL enables safe exploration and alleviates the data-inefficiency problem as it exploits prior knowledge about valid sequences of motor actions. Moreover, we provide a set of measures for evaluation of generative models such that we are able to predict the performance of the RL policy training prior to the actual training on a physical robot. We experimentally determine the characteristics of generative models that have most influence on the performance of the final policy training on two robotics tasks: shooting a hockey puck and throwing a basketball. Furthermore, we empirically demonstrate that GenRL is the only method which can safely and efficiently solve the robotics tasks compared to two state-of-the-art RL methods.

Gabriele Tiboni, Karol Arndt, Giuseppe Averta et al.

Physics simulators have shown great promise for conveniently learning reinforcement learning policies in safe, unconstrained environments. However, transferring the acquired knowledge to the real world can be challenging due to the reality gap. To this end, several methods have been recently proposed to automatically tune simulator parameters with posterior distributions given real data, for use with domain randomization at training time. These approaches have been shown to work for various robotic tasks under different settings and assumptions. Nevertheless, existing literature lacks a thorough comparison of existing adaptive domain randomization methods with respect to transfer performance and real-data efficiency. In this work, we present an open benchmark for both offline and online methods (SimOpt, BayRn, DROID, DROPO), to shed light on which are most suitable for each setting and task at hand. We found that online methods are limited by the quality of the currently learned policy for the next iteration, while offline methods may sometimes fail when replaying trajectories in simulation with open-loop commands. The code used will be released at https://github.com/gabrieletiboni/adr-benchmark.

Quentin Bouniot, Ievgen Redko, Anton Mallasto et al.

In the last decade, we have witnessed the introduction of several novel deep neural network (DNN) architectures exhibiting ever-increasing performance across diverse tasks. Explaining the upward trend of their performance, however, remains difficult as different DNN architectures of comparable depth and width -- common factors associated with their expressive power -- may exhibit a drastically different performance even when trained on the same dataset. In this paper, we introduce the concept of the non-linearity signature of DNN, the first theoretically sound solution for approximately measuring the non-linearity of deep neural networks. Built upon a score derived from closed-form optimal transport mappings, this signature provides a better understanding of the inner workings of a wide range of DNN architectures and learning paradigms, with a particular emphasis on the computer vision task. We provide extensive experimental results that highlight the practical usefulness of the proposed non-linearity signature and its potential for long-reaching implications. The code for our work is available at https://github.com/qbouniot/AffScoreDeep

Lukas Rustler, Jens Lundell, Jan Kristof Behrens et al.

Recent advancements in object shape completion have enabled impressive object reconstructions using only visual input. However, due to self-occlusion, the reconstructions have high uncertainty in the occluded object parts, which negatively impacts the performance of downstream robotic tasks such as grasping. In this work, we propose an active visuo-haptic shape completion method called Act-VH that actively computes where to touch the objects based on the reconstruction uncertainty. Act-VH reconstructs objects from point clouds and calculates the reconstruction uncertainty using IGR, a recent state-of-the-art implicit surface deep neural network. We experimentally evaluate the reconstruction accuracy of Act-VH against five baselines in simulation and in the real world. We also propose a new simulation environment for this purpose. The results show that Act-VH outperforms all baselines and that an uncertainty-driven haptic exploration policy leads to higher reconstruction accuracy than a random policy and a policy driven by Gaussian Process Implicit Surfaces. As a final experiment, we evaluate Act-VH and the best reconstruction baseline on grasping 10 novel objects. The results show that Act-VH reaches a significantly higher grasp success rate than the baseline on all objects. Together, this work opens up the door for using active visuo-haptic shape completion in more complex cluttered scenes.

David Blanco-Mulero, Gokhan Alcan, Fares J. Abu-Dakka et al.

Pre-defined manipulation primitives are widely used for cloth manipulation. However, cloth properties such as its stiffness or density can highly impact the performance of these primitives. Although existing solutions have tackled the parameterisation of pick and place locations, the effect of factors such as the velocity or trajectory of quasi-static and dynamic manipulation primitives has been neglected. Choosing appropriate values for these parameters is crucial to cope with the range of materials present in house-hold cloth objects. To address this challenge, we introduce the Quasi-Dynamic Parameterisable (QDP) method, which optimises parameters such as the motion velocity in addition to the pick and place positions of quasi-static and dynamic manipulation primitives. In this work, we leverage the framework of Sequential Reinforcement Learning to decouple sequentially the parameters that compose the primitives. To evaluate the effectiveness of the method we focus on the task of cloth unfolding with a robotic arm in simulation and real-world experiments. Our results in simulation show that by deciding the optimal parameters for the primitives the performance can improve by 20% compared to sub-optimal ones. Real-world results demonstrate the advantage of modifying the velocity and height of manipulation primitives for cloths with different mass, stiffness, shape and size. Supplementary material, videos, and code, can be found at https://sites.google.com/view/qdp-srl.

Chang Rajani, Karol Arndt, David Blanco-Mulero et al.

The co-adaptation of robots has been a long-standing research endeavour with the goal of adapting both body and behaviour of a system for a given task, inspired by the natural evolution of animals. Co-adaptation has the potential to eliminate costly manual hardware engineering as well as improve the performance of systems. The standard approach to co-adaptation is to use a reward function for optimizing behaviour and morphology. However, defining and constructing such reward functions is notoriously difficult and often a significant engineering effort. This paper introduces a new viewpoint on the co-adaptation problem, which we call co-imitation: finding a morphology and a policy that allow an imitator to closely match the behaviour of a demonstrator. To this end we propose a co-imitation methodology for adapting behaviour and morphology by matching state distributions of the demonstrator. Specifically, we focus on the challenging scenario with mismatched state- and action-spaces between both agents. We find that co-imitation increases behaviour similarity across a variety of tasks and settings, and demonstrate co-imitation by transferring human walking, jogging and kicking skills onto a simulated humanoid.

Shoaib Azam, Farzeen Munir, Ville Kyrki et al.

Learning contextual and spatial environmental representations enhances autonomous vehicle's hazard anticipation and decision-making in complex scenarios. Recent perception systems enhance spatial understanding with sensor fusion but often lack full environmental context. Humans, when driving, naturally employ neural maps that integrate various factors such as historical data, situational subtleties, and behavioral predictions of other road users to form a rich contextual understanding of their surroundings. This neural map-based comprehension is integral to making informed decisions on the road. In contrast, even with their significant advancements, autonomous systems have yet to fully harness this depth of human-like contextual understanding. Motivated by this, our work draws inspiration from human driving patterns and seeks to formalize the sensor fusion approach within an end-to-end autonomous driving framework. We introduce a framework that integrates three cameras (left, right, and center) to emulate the human field of view, coupled with top-down bird-eye-view semantic data to enhance contextual representation. The sensor data is fused and encoded using a self-attention mechanism, leading to an auto-regressive waypoint prediction module. We treat feature representation as a sequential problem, employing a vision transformer to distill the contextual interplay between sensor modalities. The efficacy of the proposed method is experimentally evaluated in both open and closed-loop settings. Our method achieves displacement error by 0.67m in open-loop settings, surpassing current methods by 6.9% on the nuScenes dataset. In closed-loop evaluations on CARLA's Town05 Long and Longest6 benchmarks, the proposed method enhances driving performance, route completion, and reduces infractions.

Francesco Verdoja, Tomasz Piotr Kucner, Ville Kyrki

Mapping people dynamics is a crucial skill for robots, because it enables them to coexist in human-inhabited environments. However, learning a model of people dynamics is a time consuming process which requires observation of large amount of people moving in an environment. Moreover, approaches for mapping dynamics are unable to transfer the learned models across environments: each model is only able to describe the dynamics of the environment it has been built in. However, the impact of architectural geometry on people's movement can be used to anticipate their patterns of dynamics, and recent work has looked into learning maps of dynamics from occupancy. So far however, approaches based on trajectories and those based on geometry have not been combined. In this work we propose a novel Bayesian approach to learn people dynamics able to combine knowledge about the environment geometry with observations from human trajectories. An occupancy-based deep prior is used to build an initial transition model without requiring any observations of pedestrian; the model is then updated when observations become available using Bayesian inference. We demonstrate the ability of our model to increase data efficiency and to generalize across real large-scale environments, which is unprecedented for maps of dynamics.

Eshagh Kargar, Ville Kyrki

Offline reinforcement learning allows training reinforcement learning models on data from live deployments. However, it is limited to choosing the best combination of behaviors present in the training data. In contrast, simulation environments attempting to replicate the live environment can be used instead of the live data, yet this approach is limited by the simulation-to-reality gap, resulting in a bias. In an attempt to get the best of both worlds, we propose a method that combines an imperfect simulation environment with data from the target environment, to train an offline reinforcement learning policy. Our experiments demonstrate that the proposed method outperforms state-of-the-art approaches CQL, MOPO, and COMBO, especially in scenarios with diverse and challenging dynamics, and demonstrates robust behavior across a variety of experimental conditions. The results highlight that using simulator-generated data can effectively enhance offline policy learning despite the sim-to-real gap, when direct interaction with the real-world is not possible.

Farzeen Munir, Shoaib Azam, Tomasz Kucner et al.

Object detection is a core component of perception systems, providing the ego vehicle with information about its surroundings to ensure safe route planning. While cameras and Lidar have significantly advanced perception systems, their performance can be limited in adverse weather conditions. In contrast, millimeter-wave technology enables radars to function effectively in such conditions. However, relying solely on radar for building a perception system doesn't fully capture the environment due to the data's sparse nature. To address this, sensor fusion strategies have been introduced. We propose a dual-branch framework to integrate radar and Lidar data for enhanced object detection. The primary branch focuses on extracting radar features, while the auxiliary branch extracts Lidar features. These are then combined using additive attention. Subsequently, the integrated features are processed through a novel Parallel Forked Structure (PFS) to manage scale variations. A region proposal head is then utilized for object detection. We evaluated the effectiveness of our proposed method on the Radiate dataset using COCO metrics. The results show that it surpasses state-of-the-art methods by $1.89\%$ and $2.61\%$ in favorable and adverse weather conditions, respectively. This underscores the value of radar-Lidar fusion in achieving precise object detection and localization, especially in challenging weather conditions.

Shivam Chaubey, Francesco Verdoja, Shankar Deka et al.

Shared control combines human intention with autonomous decision-making. At the low level, the primary goal is to maintain safety regardless of the user's input to the system. However, existing shared control methods-based on, e.g., Model Predictive Control, Control Barrier Functions, or learning-based control-often face challenges with feasibility, scalability, and mixed constraints. To address these challenges, we propose a Constraint-Aware Assistive Controller that computes control actions online while ensuring recursive feasibility, strict constraint satisfaction, and minimal deviation from the user's intent. It also accommodates a structured class of non-convex constraints common in real-world settings. We leverage Robust Controlled Invariant Sets for recursive feasibility and a Mixed-Integer Quadratic Programming formulation to handle non-convex constraints. We validate the approach through a large-scale user study with 66 participants-one of the most extensive in shared control research-using a simulated environment to assess task load, trust, and perceived control, in addition to performance. The results show consistent improvements across all these aspects without compromising safety and user intent. Additionally, a real-world experiment on a robotic manipulator demonstrates the framework's applicability under bounded disturbances, ensuring safety and collision-free operation.

Samuli Hynninen, Ville Kyrki

Manipulating open liquid containers is challenging because liquids are highly sensitive to vessel accelerations and jerks. Although spill-free liquid manipulation has been widely studied, emergency stopping under unexpected hazards has received little attention, despite the fact that abrupt braking may cause hazardous spills. This letter presents an emergency stop system for robots manipulating liquids in open containers. We formulate emergency stopping as an optimal control problem and solve it in a model predictive control framework to generate time-optimal, spill-free stopping trajectories. The method operates as a plug-and-play safety layer on top of existing slosh-free motion planning methods, enabling immediate reaction to detected hazards while accounting for nonlinear liquid dynamics. We demonstrate, through simulation and on a 7-DoF Franka Emika Panda robot, that the proposed approach achieves fast emergency stopping without spilling.

Oliver Struckmeier, Ievgen Redko, Anton Mallasto et al.

Optimal transport (OT) is a powerful geometric tool used to compare and align probability measures following the least effort principle. Despite its widespread use in machine learning (ML), OT problem still bears its computational burden, while at the same time suffering from the curse of dimensionality for measures supported on general high-dimensional spaces. In this paper, we propose to tackle these challenges using representation learning. In particular, we seek to learn an embedding space such that the samples of the two input measures become alignable in it with a simple affine mapping that can be calculated efficiently in closed-form. We then show that such approach leads to results that are comparable to solving the original OT problem when applied to the transfer learning task on which many OT baselines where previously evaluated in both homogeneous and heterogeneous DA settings. The code for our contribution is available at \url{https://github.com/Oleffa/LaOT}.

Eshagh Kargar, Ville Kyrki

This work considers the problem of learning cooperative policies in multi-agent settings with partially observable and non-stationary environments without a communication channel. We focus on improving information sharing between agents and propose a new multi-agent actor-critic method called \textit{Multi-Agent Cooperative Recurrent Proximal Policy Optimization} (MACRPO). We propose two novel ways of integrating information across agents and time in MACRPO: First, we use a recurrent layer in critic's network architecture and propose a new framework to use a meta-trajectory to train the recurrent layer. This allows the network to learn the cooperation and dynamics of interactions between agents, and also handle partial observability. Second, we propose a new advantage function that incorporates other agents' rewards and value functions. We evaluate our algorithm on three challenging multi-agent environments with continuous and discrete action spaces, Deepdrive-Zero, Multi-Walker, and Particle environment. We compare the results with several ablations and state-of-the-art multi-agent algorithms such as QMIX and MADDPG and also single-agent methods with shared parameters between agents such as IMPALA and APEX. The results show superior performance against other algorithms. The code is available online at https://github.com/kargarisaac/macrpo.

Jens Lundell, Enric Corona, Tran Nguyen Le et al.

While there exists many methods for manipulating rigid objects with parallel-jaw grippers, grasping with multi-finger robotic hands remains a quite unexplored research topic. Reasoning and planning collision-free trajectories on the additional degrees of freedom of several fingers represents an important challenge that, so far, involves computationally costly and slow processes. In this work, we present Multi-FinGAN, a fast generative multi-finger grasp sampling method that synthesizes high quality grasps directly from RGB-D images in about a second. We achieve this by training in an end-to-end fashion a coarse-to-fine model composed of a classification network that distinguishes grasp types according to a specific taxonomy and a refinement network that produces refined grasp poses and joint angles. We experimentally validate and benchmark our method against a standard grasp-sampling method on 790 grasps in simulation and 20 grasps on a real Franka Emika Panda. All experimental results using our method show consistent improvements both in terms of grasp quality metrics and grasp success rate. Remarkably, our approach is up to 20-30 times faster than the baseline, a significant improvement that opens the door to feedback-based grasp re-planning and task informative grasping. Code is available at https://irobotics.aalto.fi/multi-fingan/.

Kshitij Tiwari, Ville Kyrki, Allen Cheung et al.

With the advent of consumer-grade products for presenting an immersive virtual environment (VE), there is a growing interest in utilizing VEs for testing human navigation behavior. However, preparing a VE still requires a high level of technical expertise in computer graphics and virtual reality, posing a significant hurdle to embracing the emerging technology. To address this issue, this paper presents Delayed Feedback based Immersive Navigation Environment (DeFINE), a framework that allows for easy creation and administration of navigation tasks within customizable VEs via intuitive graphical user interfaces and simple settings files. Importantly, DeFINE has a built-in capability to provide performance feedback to participants during an experiment, a feature that is critically missing in other similar frameworks. To show the usability of DeFINE from both experimentalists' and participants' perspectives, a demonstration was made in which participants navigated to a hidden goal location with feedback that differentially weighted speed and accuracy of their responses. In addition, the participants evaluated DeFINE in terms of its ease of use, required workload, and proneness to induce cybersickness. The demonstration exemplified typical experimental manipulations DeFINE accommodates and what types of data it can collect for characterizing participants' task performance. With its out-of-the-box functionality and potential customizability due to open-source licensing, DeFINE makes VEs more accessible to many researchers.

Andrej Kruzliak, Jiri Hartvich, Shubhan P. Patni et al.

This work presents a framework for automatically extracting physical object properties, such as material composition, mass, volume, and stiffness, through robot manipulation and a database of object measurements. The framework involves exploratory action selection to maximize learning about objects on a table. A Bayesian network models conditional dependencies between object properties, incorporating prior probability distributions and uncertainty associated with measurement actions. The algorithm selects optimal exploratory actions based on expected information gain and updates object properties through Bayesian inference. Experimental evaluation demonstrates effective action selection compared to a baseline and correct termination of the experiments if there is nothing more to be learned. The algorithm proved to behave intelligently when presented with trick objects with material properties in conflict with their appearance. The robot pipeline integrates with a logging module and an online database of objects, containing over 24,000 measurements of 63 objects with different grippers. All code and data are publicly available, facilitating automatic digitization of objects and their physical properties through exploratory manipulations.

Daulet Baimukashev, Gokhan Alcan, Ville Kyrki

Inverse reinforcement learning (IRL) is an imitation learning approach to learning reward functions from expert demonstrations. Its use avoids the difficult and tedious procedure of manual reward specification while retaining the generalization power of reinforcement learning. In IRL, the reward is usually represented as a linear combination of features. In continuous state spaces, the state variables alone are not sufficiently rich to be used as features, but which features are good is not known in general. To address this issue, we propose a method that employs polynomial basis functions to form a candidate set of features, which are shown to allow the matching of statistical moments of state distributions. Feature selection is then performed for the candidates by leveraging the correlation between trajectory probabilities and feature expectations. We demonstrate the approach's effectiveness by recovering reward functions that capture expert policies across non-linear control tasks of increasing complexity. Code, data, and videos are available at https://sites.google.com/view/feature4irl.

Mingwei Deng, Ville Kyrki, Dominik Baumann

Transferring knowledge from one environment to another is an essential ability of intelligent systems. Nevertheless, when two environments are different, naively transferring all knowledge may deteriorate the performance, a phenomenon known as negative transfer. In this paper, we address this issue within the framework of multi-armed bandits from the perspective of causal inference. Specifically, we consider transfer learning in latent contextual bandits, where the actual context is hidden, but a potentially high-dimensional proxy is observable. We further consider a covariate shift in the context across environments. We show that naively transferring all knowledge for classical bandit algorithms in this setting led to negative transfer. We then leverage transportability theory from causal inference to develop algorithms that explicitly transfer effective knowledge for estimating the causal effects of interest in the target environment. Besides, we utilize variational autoencoders to approximate causal effects under the presence of a high-dimensional proxy. We test our algorithms on synthetic and semi-synthetic datasets, empirically demonstrating consistently improved learning efficiency across different proxies compared to baseline algorithms, showing the effectiveness of our causal framework in transferring knowledge.

Oskar Keurulainen, Gokhan Alcan, Ville Kyrki

Building machines capable of efficiently collaborating with humans has been a longstanding goal in artificial intelligence. Especially in the presence of uncertainties, optimal cooperation often requires that humans and artificial agents model each other's behavior and use these models to infer underlying goals, beliefs or intentions, potentially involving multiple levels of recursion. Empirical evidence for such higher-order cognition in human behavior is also provided by previous works in cognitive science, linguistics, and robotics. We advocate for a new paradigm for active learning for human feedback that utilises humans as active data sources while accounting for their higher levels of agency. In particular, we discuss how increasing level of agency results in qualitatively different forms of rational communication between an active learning system and a teacher. Additionally, we provide a practical example of active learning using a higher-order cognitive model. This is accompanied by a computational study that underscores the unique behaviors that this model produces.

Maryam Kazemi Eskeri, Thomas Wiedemann, Ville Kyrki et al.

Robots operating in human-populated environments must navigate safely and efficiently while minimizing social disruption. Achieving this requires estimating crowd movement to avoid congested areas in real-time. Traditional microscopic models struggle to scale in dense crowds due to high computational cost, while existing macroscopic crowd prediction models tend to be either overly simplistic or computationally intensive. In this work, we propose a lightweight, real-time macroscopic crowd prediction model tailored for human motion, which balances prediction accuracy and computational efficiency. Our approach simplifies both spatial and temporal processing based on the inherent characteristics of pedestrian flow, enabling robust generalization without the overhead of complex architectures. We demonstrate a 3.6 times reduction in inference time, while improving prediction accuracy by 3.1 %. Integrated into a socially aware planning framework, the model enables efficient and socially compliant robot navigation in dynamic environments. This work highlights that efficient human crowd modeling enables robots to navigate dense environments without costly computations.

Yuying Zhang, Kevin Sebastian Luck, Francesco Verdoja et al.

Mobile manipulation is a critical capability for robots operating in diverse, real-world environments. However, manipulating deformable objects and materials remains a major challenge for existing robot learning algorithms. While various benchmarks have been proposed to evaluate manipulation strategies with rigid objects, there is still a notable lack of standardized benchmarks that address mobile manipulation tasks involving deformable objects. To address this gap, we introduce MoDeSuite, the first Mobile Manipulation Deformable Object task suite, designed specifically for robot learning. MoDeSuite consists of eight distinct mobile manipulation tasks covering both elastic objects and deformable objects, each presenting a unique challenge inspired by real-world robot applications. Success in these tasks requires effective collaboration between the robot's base and manipulator, as well as the ability to exploit the deformability of the objects. To evaluate and demonstrate the use of the proposed benchmark, we train two state-of-the-art reinforcement learning algorithms and two imitation learning algorithms, highlighting the difficulties encountered and showing their performance in simulation. Furthermore, we demonstrate the practical relevance of the suite by deploying the trained policies directly into the real world with the Spot robot, showcasing the potential for sim-to-real transfer. We expect that MoDeSuite will open a novel research domain in mobile manipulation involving deformable objects. Find more details, code, and videos at https://sites.google.com/view/modesuite/home.

Chengmin Zhou, Ville Kyrki, Pasi Fränti et al.

Bayesian inference has many advantages in decision making of agents (e.g. robotics/simulative agent) over a regular data-driven black-box neural network: Data-efficiency, generalization, interpretability, and safety where these advantages benefit directly/indirectly from the uncertainty quantification of Bayesian inference. However, there are few comprehensive reviews to summarize the progress of Bayesian inference on reinforcement learning (RL) for decision making to give researchers a systematic understanding. This paper focuses on combining Bayesian inference with RL that nowadays is an important approach in agent decision making. To be exact, this paper discusses the following five topics: 1) Bayesian methods that have potential for agent decision making. First basic Bayesian methods and models (Bayesian rule, Bayesian learning, and Bayesian conjugate models) are discussed followed by variational inference, Bayesian optimization, Bayesian deep learning, Bayesian active learning, Bayesian generative models, Bayesian meta-learning, and lifelong Bayesian learning. 2) Classical combinations of Bayesian methods with model-based RL (with approximation methods), model-free RL, and inverse RL. 3) Latest combinations of potential Bayesian methods with RL. 4) Analytical comparisons of methods that combine Bayesian methods with RL with respect to data-efficiency, generalization, interpretability, and safety. 5) In-depth discussions in six complex problem variants of RL, including unknown reward, partial-observability, multi-agent, multi-task, non-linear non-Gaussian, and hierarchical RL problems and the summary of how Bayesian methods work in the data collection, data processing and policy learning stages of RL to pave the way for better agent decision-making strategies.

Kalle Kujanpää, Daulet Baimukashev, Farzeen Munir et al.

Learning to perform accurate and rich simulations of human driving behaviors from data for autonomous vehicle testing remains challenging due to human driving styles' high diversity and variance. We address this challenge by proposing a novel approach that leverages contrastive learning to extract a dictionary of driving styles from pre-existing human driving data. We discretize these styles with quantization, and the styles are used to learn a conditional diffusion policy for simulating human drivers. Our empirical evaluation confirms that the behaviors generated by our approach are both safer and more human-like than those of the machine-learning-based baseline methods. We believe this has the potential to enable higher realism and more effective techniques for evaluating and improving the performance of autonomous vehicles.

Shibei Zhu, Tran Nguyen Le, Samuel Kaski et al.

Real-time collaboration with humans poses challenges due to the different behavior patterns of humans resulting from diverse physical constraints. Existing works typically focus on learning safety constraints for collaboration, or how to divide and distribute the subtasks between the participating agents to carry out the main task. In contrast, we propose to learn a human constraints model that, in addition, considers the diverse behaviors of different human operators. We consider a type of collaboration in a shared-autonomy fashion, where both a human operator and an assistive robot act simultaneously in the same task space that affects each other's actions. The task of the assistive agent is to augment the skill of humans to perform a shared task by supporting humans as much as possible, both in terms of reducing the workload and minimizing the discomfort for the human operator. Therefore, we propose an augmentative assistant agent capable of learning and adapting to human physical constraints, aligning its actions with the ergonomic preferences and limitations of the human operator.

Andrei Aksjonov, Ville Kyrki

While artificial-intelligence-based methods suffer from lack of transparency, rule-based methods dominate in safety-critical systems. Yet, the latter cannot compete with the first ones in robustness to multiple requirements, for instance, simultaneously addressing safety, comfort, and efficiency. Hence, to benefit from both methods they must be joined in a single system. This paper proposes a decision making and control framework, which profits from advantages of both the rule- and machine-learning-based techniques while compensating for their disadvantages. The proposed method embodies two controllers operating in parallel, called Safety and Learned. A rule-based switching logic selects one of the actions transmitted from both controllers. The Safety controller is prioritized every time, when the Learned one does not meet the safety constraint, and also directly participates in the safe Learned controller training. Decision making and control in autonomous driving is chosen as the system case study, where an autonomous vehicle learns a multi-task policy to safely cross an unprotected intersection. Multiple requirements (i.e., safety, efficiency, and comfort) are set for vehicle operation. A numerical simulation is performed for the proposed framework validation, where its ability to satisfy the requirements and robustness to changing environment is successfully demonstrated.

Rituraj Kaushik, Karol Arndt, Ville Kyrki

The framework of Simulation-to-real learning, i.e, learning policies in simulation and transferring those policies to the real world is one of the most promising approaches towards data-efficient learning in robotics. However, due to the inevitable reality gap between the simulation and the real world, a policy learned in the simulation may not always generate a safe behaviour on the real robot. As a result, during adaptation of the policy in the real world, the robot may damage itself or cause harm to its surroundings. In this work, we introduce a novel learning algorithm called SafeAPT that leverages a diverse repertoire of policies evolved in the simulation and transfers the most promising safe policy to the real robot through episodic interaction. To achieve this, SafeAPT iteratively learns a probabilistic reward model as well as a safety model using real-world observations combined with simulated experiences as priors. Then, it performs Bayesian optimization on the repertoire with the reward model while maintaining the specified safety constraint using the safety model. SafeAPT allows a robot to adapt to a wide range of goals safely with the same repertoire of policies evolved in the simulation. We compare SafeAPT with several baselines, both in simulated and real robotic experiments and show that SafeAPT finds high-performance policies within a few minutes in the real world while minimizing safety violations during the interactions.

Gabriele Tiboni, Karol Arndt, Ville Kyrki

In recent years, domain randomization over dynamics parameters has gained a lot of traction as a method for sim-to-real transfer of reinforcement learning policies in robotic manipulation; however, finding optimal randomization distributions can be difficult. In this paper, we introduce DROPO, a novel method for estimating domain randomization distributions for safe sim-to-real transfer. Unlike prior work, DROPO only requires a limited, precollected offline dataset of trajectories, and explicitly models parameter uncertainty to match real data using a likelihood-based approach. We demonstrate that DROPO is capable of recovering dynamic parameter distributions in simulation and finding a distribution capable of compensating for an unmodeled phenomenon. We also evaluate the method in two zero-shot sim-to-real transfer scenarios, showing successful domain transfer and improved performance over prior methods.

Markku Suomalainen, Yiannis Karayiannidis, Ville Kyrki

In this survey, we present the current status on robots performing manipulation tasks that require varying contact with the environment, such that the robot must either implicitly or explicitly control the contact force with the environment to complete the task. Robots can perform more and more manipulation tasks that are still done by humans, and there is a growing number of publications on the topics of 1) performing tasks that always require contact and 2) mitigating uncertainty by leveraging the environment in tasks that, under perfect information, could be performed without contact. The recent trends have seen robots perform tasks earlier left for humans, such as massage, and in the classical tasks, such as peg-in-hole, there is a more efficient generalization to other similar tasks, better error tolerance, and faster planning or learning of the tasks. Thus, in this survey we cover the current stage of robots performing such tasks, starting from surveying all the different in-contact tasks robots can perform, observing how these tasks are controlled and represented, and finally presenting the learning and planning of the skills required to complete these tasks.

Neea Tuomainen, David Blanco-Mulero, Ville Kyrki

Manipulation of granular materials such as sand or rice remains an unsolved problem due to challenges such as the difficulty of defining their configuration or modeling the materials and their particles interactions. Current approaches tend to simplify the material dynamics and omit the interactions between the particles. In this paper, we propose to use a graph-based representation to model the interaction dynamics of the material and rigid bodies manipulating it. This allows the planning of manipulation trajectories to reach a desired configuration of the material. We use a graph neural network (GNN) to model the particle interactions via message-passing. To plan manipulation trajectories, we propose to minimise the Wasserstein distance between a predicted distribution of granular particles and their desired configuration. We demonstrate that the proposed method is able to pour granular materials into the desired configuration both in simulated and real scenarios.

Jouko Kinnari, Francesco Verdoja, Ville Kyrki

Localization without Global Navigation Satellite Systems (GNSS) is a critical functionality in autonomous operations of unmanned aerial vehicles (UAVs). Vision-based localization on a known map can be an effective solution, but it is burdened by two main problems: places have different appearance depending on weather and season, and the perspective discrepancy between the UAV camera image and the map make matching hard. In this work, we propose a localization solution relying on matching of UAV camera images to georeferenced orthophotos with a trained convolutional neural network model that is invariant to significant seasonal appearance difference (winter-summer) between the camera image and map. We compare the convergence speed and localization accuracy of our solution to six reference methods. The results show major improvements with respect to reference methods, especially under high seasonal variation. We finally demonstrate the ability of the method to successfully localize a real UAV, showing that the proposed method is robust to perspective changes.

Eshagh Kargar, Ville Kyrki

Driving in a complex urban environment is a difficult task that requires a complex decision policy. In order to make informed decisions, one needs to gain an understanding of the long-range context and the importance of other vehicles. In this work, we propose to use Vision Transformer (ViT) to learn a driving policy in urban settings with birds-eye-view (BEV) input images. The ViT network learns the global context of the scene more effectively than with earlier proposed Convolutional Neural Networks (ConvNets). Furthermore, ViT's attention mechanism helps to learn an attention map for the scene which allows the ego car to determine which surrounding cars are important to its next decision. We demonstrate that a DQN agent with a ViT backbone outperforms baseline algorithms with ConvNet backbones pre-trained in various ways. In particular, the proposed method helps reinforcement learning algorithms to learn faster, with increased performance and less data than baselines.

Tran Nguyen Le, Jens Lundell, Fares J. Abu-Dakka et al.

Grasp synthesis for 3D deformable objects remains a little-explored topic, most works aiming to minimize deformations. However, deformations are not necessarily harmful -- humans are, for example, able to exploit deformations to generate new potential grasps. How to achieve that on a robot is though an open question. This paper proposes an approach that uses object stiffness information in addition to depth images for synthesizing high-quality grasps. We achieve this by incorporating object stiffness as an additional input to a state-of-the-art deep grasp planning network. We also curate a new synthetic dataset of grasps on objects of varying stiffness using the Isaac Gym simulator for training the network. We experimentally validate and compare our proposed approach against the case where we do not incorporate object stiffness on a total of 2800 grasps in simulation and 420 grasps on a real Franka Emika Panda. The experimental results show significant improvement in grasp success rate using the proposed approach on a wide range of objects with varying shapes, sizes, and stiffness. Furthermore, we demonstrate that the approach can generate different grasping strategies for different stiffness values, such as pinching for soft objects and caging for hard objects. Together, the results clearly show the value of incorporating stiffness information when grasping objects of varying stiffness.

Julius Hietala, David Blanco-Mulero, Gokhan Alcan et al.

Robotic manipulation of cloth is a challenging task due to the high dimensionality of the configuration space and the complexity of dynamics affected by various material properties. The effect of complex dynamics is even more pronounced in dynamic folding, for example, when a square piece of fabric is folded in two by a single manipulator. To account for the complexity and uncertainties, feedback of the cloth state using e.g. vision is typically needed. However, construction of visual feedback policies for dynamic cloth folding is an open problem. In this paper, we present a solution that learns policies in simulation using Reinforcement Learning (RL) and transfers the learned policies directly to the real world. In addition, to learn a single policy that manipulates multiple materials, we randomize the material properties in simulation. We evaluate the contributions of visual feedback and material randomization in real-world experiments. The experimental results demonstrate that the proposed solution can fold successfully different fabric types using dynamic manipulation in the real world. Code, data, and videos are available at https://sites.google.com/view/dynamic-cloth-folding

Tran Nguyen Le, Jens Lundell, Fares J. Abu-Dakka et al.

Grasping deformable objects is not well researched due to the complexity in modelling and simulating the dynamic behavior of such objects. However, with the rapid development of physics-based simulators that support soft bodies, the research gap between rigid and deformable objects is getting smaller. To leverage the capability of such simulators and to challenge the assumption that has guided robotic grasping research so far, i.e., object rigidity, we proposed a deep-learning based approach that generates stiffness-dependent grasps. Our network is trained on purely synthetic data generated from a physics-based simulator. The same simulator is also used to evaluate the trained network. The results show improvement in terms of grasp ranking and grasp success rate. Furthermore, our network can adapt the grasps based on the stiffness. We are currently validating the proposed approach on a larger test dataset in simulation and on a physical robot.

David Blanco-Mulero, Markus Heinonen, Ville Kyrki

Graph Gaussian Processes (GGPs) provide a data-efficient solution on graph structured domains. Existing approaches have focused on static structures, whereas many real graph data represent a dynamic structure, limiting the applications of GGPs. To overcome this we propose evolving-Graph Gaussian Processes (e-GGPs). The proposed method is capable of learning the transition function of graph vertices over time with a neighbourhood kernel to model the connectivity and interaction changes between vertices. We assess the performance of our method on time-series regression problems where graphs evolve over time. We demonstrate the benefits of e-GGPs over static graph Gaussian Process approaches.

Anton Mallasto, Karol Arndt, Markus Heinonen et al.

Sample-efficient domain adaptation is an open problem in robotics. In this paper, we present affine transport -- a variant of optimal transport, which models the mapping between state transition distributions between the source and target domains with an affine transformation. First, we derive the affine transport framework; then, we extend the basic framework with Procrustes alignment to model arbitrary affine transformations. We evaluate the method in a number of OpenAI Gym sim-to-sim experiments with simulation environments, as well as on a sim-to-real domain adaptation task of a robot hitting a hockeypuck such that it slides and stops at a target position. In each experiment, we evaluate the results when transferring between each pair of dynamics domains. The results show that affine transport can significantly reduce the model adaptation error in comparison to using the original, non-adapted dynamics model.

Jiyo Palatti, Andrei Aksjonov, Gokhan Alcan et al.

Overtaking is one of the most challenging tasks in driving, and the current solutions to autonomous overtaking are limited to simple and static scenarios. In this paper, we present a method for behaviour and trajectory planning for safe autonomous overtaking. The proposed method optimizes the trajectory by simultaneously enforcing safety and minimizing intrusion onto the adjacent lane. Furthermore, the method allows the overtaking to be aborted, enabling the autonomous vehicle to merge back in the lane, if safety is compromised, because of e.g. traffic in opposing direction appearing during the maneuver execution. A finite state machine is used to select an appropriate maneuver at each time, and a combination of safe and reachable sets is used to iteratively generate intermediate reference targets based on the current maneuver. A nonlinear model predictive controller then plans dynamically feasible and collision-free trajectories to these intermediate reference targets. Simulation experiments demonstrate that the combination of intermediate reference generation and model predictive control is able to handle multiple behaviors, including following a lead vehicle, overtaking and aborting the overtake, within a single framework.

Eshagh Kargar, Ville Kyrki

Driving in a dynamic, multi-agent, and complex urban environment is a difficult task requiring a complex decision-making policy. The learning of such a policy requires a state representation that can encode the entire environment. Mid-level representations that encode a vehicle's environment as images have become a popular choice. Still, they are quite high-dimensional, limiting their use in data-hungry approaches such as reinforcement learning. In this article, we propose to learn a low-dimensional and rich latent representation of the environment by leveraging the knowledge of relevant semantic factors. To do this, we train an encoder-decoder deep neural network to predict multiple application-relevant factors such as the trajectories of other agents and the ego car. Furthermore, we propose a hazard signal based on other vehicles' future trajectories and the planned route which is used in conjunction with the learned latent representation as input to a down-stream policy. We demonstrate that using the multi-head encoder-decoder neural network results in a more informative representation than a standard single-head model. In particular, the proposed representation learning and the hazard signal help reinforcement learning to learn faster, with increased performance and less data than baseline methods.

Jouko Kinnari, Francesco Verdoja, Ville Kyrki

Localization of low-cost Unmanned Aerial Vehicles (UAVs) often relies on Global Navigation Satellite Systems (GNSS). GNSS are susceptible to both natural disruptions to radio signal and intentional jamming and spoofing by an adversary. A typical way to provide georeferenced localization without GNSS for small UAVs is to have a downward-facing camera and match camera images to a map. The downward-facing camera adds cost, size, and weight to the UAV platform and the orientation limits its usability for other purposes. In this work, we propose a Monte-Carlo localization method for georeferenced localization of an UAV requiring no infrastructure using only inertial measurements, a camera facing an arbitrary direction, and an orthoimage map. We perform orthorectification of the UAV image, relying on a local planarity assumption of the environment, relaxing the requirement of downward-pointing camera. We propose a measure of goodness for the matching score of an orthorectified UAV image and a map. We demonstrate that the system is able to localize globally an UAV with modest requirements for initialization and map resolution.

Seppo Sierla, Ville Kyrki, Pekka Aarnio et al.

This paper proposes a new concept in which a digital twin derived from a digital product description will automatically perform assembly planning and orchestrate the production resources in a manufacturing cell. Thus the manufacturing cell has generic services with minimal assumptions about what kind of product will be assembled, while the digital product description is designed collaboratively between the designer at an OEM and automated services at potential manufacturers. This has several advantages. Firstly, the resulting versatile manufacturing facility can handle a broad variety of products with minimal or no reconfiguration effort, so it can cost-effectively offer its services to a large number of OEMs. Secondly, a solution is presented to the problem of performing concurrent product design and assembly planning over the organizational boundary. Thirdly, the product design at the OEM is not constrained to the capabilities of specific manufacturing facilities. The concept is presented in general terms in UML and an implementation is provided in a 3D simulation environment using Automation Markup Language for digital product descriptions. Finally, two case studies are presented and applications in a real industrial context are discussed.

Krishnananda Prabhu Sivananda, Francesco Verdoja, Ville Kyrki

While 2D occupancy maps commonly used in mobile robotics enable safe navigation in indoor environments, in order for robots to understand and interact with their environment and its inhabitants representing 3D geometry and semantic environment information is required. Semantic information is crucial in effective interpretation of the meanings humans attribute to different parts of a space, while 3D geometry is important for safety and high-level understanding. We propose a pipeline that can generate a multi-layer representation of indoor environments for robotic applications. The proposed representation includes 3D metric-semantic layers, a 2D occupancy layer, and an object instance layer where known objects are replaced with an approximate model obtained through a novel model-matching approach. The metric-semantic layer and the object instance layer are combined to form an augmented representation of the environment. Experiments show that the proposed shape matching method outperforms a state-of-the-art deep learning method when tasked to complete unseen parts of objects in the scene. The pipeline performance translates well from simulation to real world as shown by F1-score analysis, with semantic segmentation accuracy using Mask R-CNN acting as the major bottleneck. Finally, we also demonstrate on a real robotic platform how the multi-layer map can be used to improve navigation safety.

Karol Arndt, Oliver Struckmeier, Ville Kyrki

Domain adaptation is a common problem in robotics, with applications such as transferring policies from simulation to real world and lifelong learning. Performing such adaptation, however, requires informative data about the environment to be available during the adaptation. In this paper, we present domain curiosity -- a method of training exploratory policies that are explicitly optimized to provide data that allows a model to learn about the unknown aspects of the environment. In contrast to most curiosity methods, our approach explicitly rewards learning, which makes it robust to environment noise without sacrificing its ability to learn. We evaluate the proposed method by comparing how much a model can learn about environment dynamics given data collected by the proposed approach, compared to standard curious and random policies. The evaluation is performed using a toy environment, two simulated robot setups, and on a real-world haptic exploration task. The results show that the proposed method allows data-efficient and accurate estimation of dynamics.

Jens Lundell, Francesco Verdoja, Ville Kyrki

Recent advances in multi-fingered robotic grasping have enabled fast 6-Degrees-Of-Freedom (DOF) single object grasping. Multi-finger grasping in cluttered scenes, on the other hand, remains mostly unexplored due to the added difficulty of reasoning over obstacles which greatly increases the computational time to generate high-quality collision-free grasps. In this work we address such limitations by introducing DDGC, a fast generative multi-finger grasp sampling method that can generate high quality grasps in cluttered scenes from a single RGB-D image. DDGC is built as a network that encodes scene information to produce coarse-to-fine collision-free grasp poses and configurations. We experimentally benchmark DDGC against the simulated-annealing planner in GraspIt! on 1200 simulated cluttered scenes and 7 real world scenes. The results show that DDGC outperforms the baseline on synthesizing high-quality grasps and removing clutter while being 5 times faster. This, in turn, opens up the door for using multi-finger grasps in practical applications which has so far been limited due to the excessive computation time needed by other methods.

Oliver Struckmeier, Kshitij Tiwari, Ville Kyrki

The slowness principle is a concept inspired by the visual cortex of the brain. It postulates that the underlying generative factors of a quickly varying sensory signal change on a slower time scale. Unsupervised learning of intermediate representations utilizing abundant unlabeled sensory data can be leveraged to perform data-efficient supervised downstream regression. In this paper, we propose a general formulation of slowness for unsupervised representation learning adding a slowness regularization term to the estimate lower bound of the beta-VAE to encourage temporal similarity in observation and latent space. Within this framework we compare existing slowness regularization terms such as the L1 and L2 loss used in existing end-to-end methods, the SlowVAE and propose a new term based on Brownian motion. We empirically evaluate these slowness regularization terms with respect to their downstream task performance and data efficiency. We find that slow representations lead to equal or better downstream task performance and data efficiency in different experiment domains when compared to representations without slowness regularization. Finally, we discuss how the Frechet Inception Distance (FID), traditionally used to determine the generative capabilities of GANs, can serve as a measure to predict the performance of pre-trained Autoencoder model in a supervised downstream task and accelerate hyperparameter search.

Gokhan Alcan, Ville Kyrki

Safe operation of systems such as robots requires them to plan and execute trajectories subject to safety constraints. When those systems are subject to uncertainties in their dynamics, it is challenging to ensure that the constraints are not violated. In this paper, we propose Safe-CDDP, a safe trajectory optimization and control approach for systems under additive uncertainties and non-linear safety constraints based on constrained differential dynamic programming (DDP). The safety of the robot during its motion is formulated as chance constraints with user-chosen probabilities of constraint satisfaction. The chance constraints are transformed into deterministic ones in DDP formulation by constraint tightening. To avoid over-conservatism during constraint tightening, linear control gains of the feedback policy derived from the constrained DDP are used in the approximation of closed-loop uncertainty propagation in prediction. The proposed algorithm is empirically evaluated on three different robot dynamics with up to 12 degrees of freedom in simulation. The computational feasibility and applicability of the approach are demonstrated with a physical hardware implementation.

Joni Pajarinen, Jens Lundell, Ville Kyrki

Manipulating unknown objects in a cluttered environment is difficult because segmentation of the scene into objects, that is, object composition is uncertain. Due to this uncertainty, earlier work has concentrated on either identifying the "best" object composition and deciding on manipulation actions accordingly, or, tried to greedily gather information about the "best" object composition. Contrary to earlier work, we 1) utilize different possible object compositions in planning, 2) take advantage of object composition information provided by robot actions, 3) take into account the effect of different competing object hypotheses on the actual task to be performed. We cast the manipulation planning problem as a partially observable Markov decision process (POMDP) which plans over possible hypotheses of object compositions. The POMDP model chooses the action that maximizes the long-term expected task specific utility, and while doing so, considers the value of informative actions and the effect of different object hypotheses on the completion of the task. In simulations and in experiments with an RGB-D sensor, a Kinova Jaco and a Franka Emika Panda robot arm, a probabilistic approach outperforms an approach that only considers the most likely object composition and long term planning outperforms greedy decision making.

Karol Arndt, Ali Ghadirzadeh, Murtaza Hazara et al.

Few-shot adaptation is a challenging problem in the context of simulation-to-real transfer in robotics, requiring safe and informative data collection. In physical systems, additional challenge may be posed by domain noise, which is present in virtually all real-world applications. In this paper, we propose to perform few-shot adaptation of dynamics models in noisy conditions using an uncertainty-aware Kalman filter-based neural network architecture. We show that the proposed method, which explicitly addresses domain noise, improves few-shot adaptation error over a blackbox adaptation LSTM baseline, and over a model-free on-policy reinforcement learning approach, which tries to learn an adaptable and informative policy at the same time. The proposed method also allows for system analysis by analyzing hidden states of the model during and after adaptation.

Tran Nguyen Le, Francesco Verdoja, Fares J. Abu-Dakka et al.

Accurately modeling local surface properties of objects is crucial to many robotic applications, from grasping to material recognition. Surface properties like friction are however difficult to estimate, as visual observation of the object does not convey enough information over these properties. In contrast, haptic exploration is time consuming as it only provides information relevant to the explored parts of the object. In this work, we propose a joint visuo-haptic object model that enables the estimation of surface friction coefficient over an entire object by exploiting the correlation of visual and haptic information, together with a limited haptic exploration by a robotic arm. We demonstrate the validity of the proposed method by showing its ability to estimate varying friction coefficients on a range of real multi-material objects. Furthermore, we illustrate how the estimated friction coefficients can improve grasping success rate by guiding a grasp planner toward high friction areas.