Below is a list of all relevant publications authored by Robotics Forum members.
- Showing results for:
- Reset all filters
Conference paperCursi F, Bai W, Kormushev P, 2021,
Robot modelling is an essential part to properlyunderstand how a robotic system moves and how to controlit. The kinematic model of a robot is usually obtained byusing Denavit-Hartenberg convention, which relies on a set ofparameters to describe the end-effector pose in a Cartesianspace. These parameters are assigned based on geometricalconsiderations of the robotic structure, however, the assignedvalues may be inaccurate. The purpose of robot kinematiccalibration is therefore to find optimal parameters whichimprove the accuracy of the robot model. In this work wepresent Kalibrot, an open source Matlab package for robotkinematic calibration. Kalibrot has been designed to simplifyrobot calibration and easily assess the calibration results. Besidecomputing the optimal parameters, Kalibrot provides a visualization layer showing the values of the calibrated parameters,what parameters can be identified, and the calibrated roboticstructure. The capabilities of the package are here shownthrough simulated and real world experiments.
Conference paperLa Barbera V, Pardo F, Tassa Y, et al., 2021,
OstrichRL: a musculoskeletal ostrich simulation to study bio-mechanical locomotion, NeurIPS 2021
Muscle-actuated control is a research topic of interest spanning different fields, inparticular biomechanics, robotics and graphics. This type of control is particularlychallenging because models are often overactuated, and dynamics are delayed andnon-linear. It is however a very well tested and tuned actuation model that hasundergone millions of years of evolution and that involves interesting propertiesexploiting passive forces of muscle-tendon units and efficient energy storage andrelease. To facilitate research on muscle-actuated simulation, we release a 3Dmusculoskeletal simulation of an ostrich based on the MuJoCo simulator. Ostrichesare one of the fastest bipeds on earth and are therefore an excellent model forstudying muscle-actuated bipedal locomotion. The model is based on CT scans anddissections used to gather actual muscle data such as insertion sites, lengths andpennation angles. Along with this model, we also provide a set of reinforcementlearning tasks, including reference motion tracking and a reaching task with theneck. The reference motion data are based on motion capture clips of variousbehaviors which we pre-processed and adapted to our model. This paper describeshow the model was built and iteratively improved using the tasks. We evaluate theaccuracy of the muscle actuation patterns by comparing them to experimentallycollected electromyographic data from locomoting birds. We believe that this workcan be a useful bridge between the biomechanics, reinforcement learning, graphicsand robotics communities, by providing a fast and easy to use simulation.
Conference paperWang K, Saputra RP, Foster JP, et al., 2021,
Improved energy efficiency via parallel elastic elements for the straight-legged vertically-compliant robot SLIDER, Japan, 24th International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines, Publisher: Springer, Pages: 129-140
Most state-of-the-art bipedal robots are designed to be anthropomorphic, and therefore possess articulated legs with knees. Whilstthis facilitates smoother, human-like locomotion, there are implementation issues that make walking with straight legs difficult. Many robotshave to move with a constant bend in the legs to avoid a singularityoccurring at the knee joints. The actuators must constantly work tomaintain this stance, which can result in the negation of energy-savingtechniques employed. Furthermore, vertical compliance disappears whenthe leg is straight and the robot undergoes high-energy loss events such asimpacts from running and jumping, as the impact force travels throughthe fully extended joints to the hips. In this paper, we attempt to improve energy efficiency in a simple yet effective way: attaching bungeecords as elastic elements in parallel to the legs of a novel, knee-less bipedrobot SLIDER, and show that the robot’s prismatic hip joints preservevertical compliance despite the legs being constantly straight. Due tothe nonlinear dynamics of the bungee cords and various sources of friction, Bayesian Optimization is utilized to find the optimals configurationof bungee cords that achieves the largest reduction in energy consumption. The optimal solution found saves 15% of the energy consumptioncompared to the robot configuration without parallel elastic elements.Additional Video: https://youtu.be/ZTaG9−Dz8A
Conference paperCully A, 2021,
Multi-Emitter MAP-Elites: Improving quality, diversity and convergence speed with heterogeneous sets of emitters, Genetic and Evolutionary Computation Conference (GECCO), Publisher: ACM, Pages: 84-92
Quality-Diversity (QD) optimisation is a new family of learning algorithmsthat aims at generating collections of diverse and high-performing solutions.Among those algorithms, MAP-Elites is a simple yet powerful approach that hasshown promising results in numerous applications. In this paper, we introduce anovel algorithm named Multi-Emitter MAP-Elites (ME-MAP-Elites) that improvesthe quality, diversity and convergence speed of MAP-Elites. It is based on therecently introduced concept of emitters, which are used to drive thealgorithm's exploration according to predefined heuristics. ME-MAP-Elitesleverages the diversity of a heterogeneous set of emitters, in which eachemitter type is designed to improve differently the optimisation process.Moreover, a bandit algorithm is used to dynamically find the best emitter setdepending on the current situation. We evaluate the performance ofME-MAP-Elites on six tasks, ranging from standard optimisation problems (in 100dimensions) to complex locomotion tasks in robotics. Our comparisons againstMAP-Elites and existing approaches using emitters show that ME-MAP-Elites isfaster at providing collections of solutions that are significantly morediverse and higher performing. Moreover, in the rare cases where no fruitfulsynergy can be found between the different emitters, ME-MAP-Elites isequivalent to the best of the compared algorithms.
Conference paperRakicevic N, Cully A, Kormushev P, 2021,
Neuroevolution is an alternative to gradient-based optimisation that has the potential to avoid local minima and allows parallelisation. The main limiting factor is that usually it does not scale well with parameter space dimensionality. Inspired by recent work examining neural network intrinsic dimension and loss landscapes, we hypothesise that there exists a low-dimensional manifold, embedded in the policy network parameter space, around which a high-density of diverse and useful policies are located. This paper proposes a novel method for diversity-based policy search via Neuroevolution, that leverages learned representations of the policy network parameters, by performing policy search in this learned representation space. Our method relies on the Quality-Diversity (QD) framework which provides a principled approach to policy search, and maintains a collection of diverse policies, used as a dataset for learning policy representations. Further, we use the Jacobian of the inverse-mapping function to guide the search in the representation space. This ensures that the generated samples remain in the high-density regions, after mapping back to the original space. Finally, we evaluate our contributions on four continuous-control tasks in simulated environments, and compare to diversity-based baselines.
Journal articleSaputra RP, Rakicevic N, Kuder I, et al., 2021,
ResQbot 2.0: an improved design of a mobile rescue robot with an inflatable neck securing device for safe casualty extraction, Applied Sciences, Vol: 11, Pages: 1-18, ISSN: 2076-3417
Despite the fact that a large number of research studies have been conducted in the field of searchand rescue robotics, significantly little attention has been given to the development of rescue robotscapable of performing physical rescue interventions, including loading and transporting victims toa safe zone—i.e. casualty extraction tasks. The aim of this study is to develop a mobile rescue robotthat could assist first responders when saving casualties from a danger area by performing a casualty extraction procedure, whilst ensuring that no additional injury is caused by the operation andno additional lives are put at risk. In this paper, we present a novel design of ResQbot 2.0—a mobilerescue robot designed for performing the casualty extraction task. This robot is a stretcher-type casualty extraction robot, which is a significantly improved version of the initial proof-of-concept prototype, ResQbot (retrospectively referred to as ResQbot 1.0), that has been developed in our previous work. The proposed designs and development of the mechanical system of ResQbot 2.0, as wellas the method for safely loading a full body casualty onto the robot’s ‘stretcher bed’, are describedin detail based on the conducted literature review, evaluation of our previous work and feedbackprovided by medical professionals. To verify the proposed design and the casualty extraction procedure, we perform simulation experiments in Gazebo physics engine simulator. The simulationresults demonstrate the capability of ResQbot 2.0 to successfully carry out safe casualty extractions
Conference paperFrazelle C, Walker I, AlAttar A, et al., 2021,
Continuum robots have strong potential for application in Space environments. However, their modeling is challenging in comparison with traditional rigid-link robots. The Kinematic-Model-Free (KMF) robot control method has been shown to be extremely effective in permitting a rigid-link robot to learn approximations of local kinematics and dynamics (“kinodynamics”) at various points in the robot's task space. These approximations enable the robot to follow various trajectories and even adapt to changes in the robot's kinematic structure. In this paper, we present the adaptation of the KMF method to a three-section, nine degrees-of-freedom continuum manipulator for both planar and spatial task spaces. Using only an external 3D camera, we show that the KMF method allows the continuum robot to converge to various desired set points in the robot's task space, avoiding the complexities inherent in solving this problem using traditional inverse kinematics. The success of the method shows that a continuum robot can “learn” enough information from an external camera to reach and track desired points and trajectories, without needing knowledge of exact shape or position of the robot. We similarly apply the method in a simulated example of a continuum robot performing an inspection task on board the ISS.
Journal articleAlAttar A, Cursi F, Kormushev P, 2021,
Kinematic-model-free redundancy resolution using multi-point tracking and control for robot manipulation, Applied Sciences, Vol: 11, Pages: 1-15, ISSN: 2076-3417
Abstract: Robots have been predominantly controlled using conventional control methods that require prior knowledge of the robots’ kinematic and dynamic models. These controllers can be challenging to tune and cannot directly adapt to changes in kinematic structure or dynamic properties. On the other hand, model-learning controllers can overcome such challenges.Our recently proposed model-learning orientation controller has shown promising ability to simul6 taneously control a three-degrees-of-freedom robot manipulator’s end-effector pose. However, this controller does not perform optimally with robots of higher degrees-of-freedom nor does it resolve redundancies. The research presented in this paper extends the state-of-the-art kinematic9 model-free controller to perform pose control of hyper-redundant robot manipulators and resolve redundancies by tracking and controlling multiple points along the robot’s serial chain. The results show that with more control points, the controller is able to reach desired poses in fewer steps, yielding an improvement of up to 66%, and capable of achieving complex configurations. The algorithm was validated by running the simulation 100 times and it was found that 82% of the times the robot successfully reached the desired target pose within 150 steps.
Conference paperTavakoli A, Fatemi M, Kormushev P, 2021,
Learning to represent action values as a hypergraph on the action vertices, Vienna, Austria, International Conference on Learning Representations
Action-value estimation is a critical component of many reinforcement learning(RL) methods whereby sample complexity relies heavily on how fast a good estimator for action value can be learned. By viewing this problem through the lens ofrepresentation learning, good representations of both state and action can facilitateaction-value estimation. While advances in deep learning have seamlessly drivenprogress in learning state representations, given the specificity of the notion ofagency to RL, little attention has been paid to learning action representations. Weconjecture that leveraging the combinatorial structure of multi-dimensional actionspaces is a key ingredient for learning good representations of action. To test this,we set forth the action hypergraph networks framework—a class of functions forlearning action representations in multi-dimensional discrete action spaces with astructural inductive bias. Using this framework we realise an agent class basedon a combination with deep Q-networks, which we dub hypergraph Q-networks.We show the effectiveness of our approach on a myriad of domains: illustrativeprediction problems under minimal confounding effects, Atari 2600 games, anddiscretised physical control benchmarks.
Journal articleRussell F, Takeda Y, Kormushev P, et al., 2021,
Stiffness modulation in walking is critical to maintain static/dynamic stability as well as minimize energy consumption and impact damage. However, optimal, or even functional, stiffness parameterization remains unresolved in legged robotics.We introduce an architecture for stiffness control utilizing a bioinspired robotic limb consisting of a condylar knee joint and leg with antagonistic actuation. The joint replicates elastic ligaments of the human knee providing tuneable compliance for walking. It further locks out at maximum extension, providing stability when standing. Compliance and friction losses between joint surfaces are derived as a function of ligament stiffness and length. Experimental studies validate utility through quantification of: 1) hip perturbation response; 2) payload capacity; and 3) static stiffness of the leg mechanism.Results prove initiation and compliance at lock out can be modulated independently of friction loss by changing ligament elasticity. Furthermore, increasing co-contraction or decreasing joint angle enables increased leg stiffness, which establishes co-contraction is counterbalanced by decreased payload.Findings have direct application in legged robots and transfemoral prosthetic knees, where biorobotic design could reduce energy expense while improving efficiency and stability. Future targeted impact involves increasing power/weight ratios in walking robots and artificial limbs for increased efficiency and precision in walking control.
Journal articleCursi F, Modugno V, Lanari L, et al., 2021,
Bayesian neural network modeling and hierarchical MPC for a tendon-driven surgical robot with uncertainty minimization, IEEE Robotics and Automation Letters, Vol: 6, Pages: 2642-2649, ISSN: 2377-3766
In order to guarantee precision and safety in robotic surgery, accurate models of the robot and proper control strategies are needed. Bayesian Neural Networks (BNN) are capable of learning complex models and provide information about the uncertainties of the learned system. Model Predictive Control (MPC) is a reliable control strategy to ensure optimality and satisfaction of safety constraints. In this work we propose the use of BNN to build the highly nonlinear kinematic and dynamic models of a tendon-driven surgical robot, and exploit the information about the epistemic uncertainties by means of a Hierarchical MPC (Hi-MPC) control strategy. Simulation and real world experiments show that the method is capable of ensuring accurate tip positioning, while satisfying imposed safety bounds on the kinematics and dynamics of the robot.
Journal articleSaputra RP, Rakicevic N, Chappell D, et al., 2021,
In recent years, several robots have been developed and deployed to perform casualty extraction tasks. However, the majority of these robots are overly complex, and require teleoperation via either a skilled operator or a specialised device, and often the operator must be present at the scene to navigate safely around the casualty. Instead, improving the autonomy of such robots can reduce the reliance on expert operators and potentially unstable communication systems, while still extracting the casualty in a safe manner. There are several stages in the casualty extraction procedure, from navigating to the location of the emergency, safely approaching and loading the casualty, to finally navigating back to the medical assistance location. In this paper, we propose a Hierarchical Decomposed-Objective based Model Predictive Control (HiDO-MPC) method for safely approaching and manoeuvring around the casualty. We implement this controller on ResQbot — a proof-of-concept mobile rescue robot we previously developed — capable of safely rescuing an injured person lying on the ground, i.e. performing the casualty extraction procedure. HiDO-MPC achieves the desired casualty extraction behaviour by decomposing the main objective into multiple sub-objectives with a hierarchical structure. At every time step, the controller evaluates this hierarchical decomposed objective and generates the optimal control decision. We have conducted a number of experiments both in simulation and using the real robot to evaluate the proposed method’s performance, and compare it with baseline approaches. The results demonstrate that the proposed control strategy gives significantly better results than baseline approaches in terms of accuracy, robustness, and execution time, when applied to casualty extraction scenarios.
Conference paperMurai R, Saeedi S, Kelly P, 2021,
Focal-plane Sensor-processor (FPSP) is a next-generation camera technology which enables every pixel on the sensor chip to perform computation in parallel, on the focal plane where the light intensity is captured. SCAMP-5 is a general-purpose FPSP used in this work and it carries out computations in the analog domain before analog to digital conversion. By extracting features from the image on the focal plane, data which is digitised and transferred is reduced. As a consequence, SCAMP-5 offers a high frame rate while maintaining low energy consumption. Here, we present BITVO, which is the first 6-Degrees of Freedom visual odometry algorithm which utilises the FPSP. Our entire system operates at 300 FPS in a natural environment, using binary edges and corner features detected by the SCAMP-5.
Conference paperJohns E, Garcia-Hernando G, Kim T-K, 2020,
Dexterous manipulation of objects in virtual environments with our bare hands, by using only a depth sensor and a state-of-the-art 3D hand pose estimator (HPE), is challenging. While virtual environments are ruled by physics, e.g. object weights and surface frictions, the absence of force feedback makes the task challenging, as even slight inaccuracies on finger tips or contact points from HPE may make the interactions fail. Prior arts simply generate contact forces in the direction of the fingers' closures, when finger joints penetrate virtual objects. Although useful for simple grasping scenarios, they cannot be applied to dexterous manipulations such as inhand manipulation. Existing reinforcement learning (RL) and imitation learning (IL) approaches train agents that learn skills by using task-specific rewards, without considering any online user input. In this work, we propose to learn a model that maps noisy input hand poses to target virtual poses, which introduces the needed contacts to accomplish the tasks on a physics simulator. The agent is trained in a residual setting by using a model-free hybrid RL+IL approach. A 3D hand pose estimation reward is introduced leading to an improvement on HPE accuracy when the physics-guided corrected target poses are remapped to the input space. As the model corrects HPE errors by applying minor but crucial joint displacements for contacts, this helps to keep the generated motion visually close to the user input. Since HPE sequences performing successful virtual interactions do not exist, a data generation scheme to train and evaluate the system is proposed. We test our framework in two applications that use hand pose estimates for dexterous manipulations: hand-object interactions in VR and hand-object motion reconstruction in-the-wild. Experiments show that the proposed method outperforms various RL/IL baselines and the simple prior art of enforcing hand closure, both in task success and hand pose accuracy.
Conference paperValassakis P, Ding Z, Johns E, 2021,
Zero-shot sim-to-real transfer of tasks with complex dynamics is a highly challenging and unsolved problem. A number of solutions have been proposed in recent years, but we have found that many works do not present a thorough evaluation in the real world, or underplay the significant engineering effort and task-specific fine tuning that is required to achieve the published results. In this paper, we dive deeper into the sim-to-real transfer challenge, investigate why this issuch a difficult problem, and present objective evaluations of anumber of transfer methods across a range of real-world tasks.Surprisingly, we found that a method which simply injects random forces into the simulation performs just as well as more complex methods, such as those which randomise the simulator's dynamics parameters
Conference paperCursi F, Modugno V, Kormushev P, 2021,
Model predictive control for a tendon-driven surgical robot with safety constraints in kinematics and dynamics, Las Vegas, USA, International Conference on Intelligence Robots and Systems (IROS), Pages: 7653-7660
In fields such as minimally invasive surgery, effective control strategies are needed to guarantee safety andaccuracy of the surgical task. Mechanical designs and actuationschemes have inevitable limitations such as backlash and jointlimits. Moreover, surgical robots need to operate in narrowpathways, which may give rise to additional environmentalconstraints. Therefore, the control strategies must be capableof satisfying the desired motion trajectories and the imposedconstraints. Model Predictive Control (MPC) has proven effective for this purpose, allowing to solve an optimal problem bytaking into consideration the evolution of the system states, costfunction, and constraints over time. The high nonlinearities intendon-driven systems, adopted in many surgical robots, are difficult to be modelled analytically. In this work, we use a modellearning approach for the dynamics of tendon-driven robots.The dynamic model is then employed to impose constraintson the torques of the robot under consideration and solve anoptimal constrained control problem for trajectory trackingby using MPC. To assess the capabilities of the proposedframework, both simulated and real world experiments havebeen conducted
Conference paperRakicevic N, Cully A, Kormushev P, 2020,
Policy manifold search for improving diversity-based neuroevolution, Publisher: arXiv
Diversity-based approaches have recently gained popularity as an alternativeparadigm to performance-based policy search. A popular approach from thisfamily, Quality-Diversity (QD), maintains a collection of high-performingpolicies separated in the diversity-metric space, defined based on policies'rollout behaviours. When policies are parameterised as neural networks, i.e.Neuroevolution, QD tends to not scale well with parameter space dimensionality.Our hypothesis is that there exists a low-dimensional manifold embedded in thepolicy parameter space, containing a high density of diverse and feasiblepolicies. We propose a novel approach to diversity-based policy search viaNeuroevolution, that leverages learned latent representations of the policyparameters which capture the local structure of the data. Our approachiteratively collects policies according to the QD framework, in order to (i)build a collection of diverse policies, (ii) use it to learn a latentrepresentation of the policy parameters, (iii) perform policy search in thelearned latent space. We use the Jacobian of the inverse transformation(i.e.reconstruction function) to guide the search in the latent space. Thisensures that the generated samples remain in the high-density regions of theoriginal space, after reconstruction. We evaluate our contributions on threecontinuous control tasks in simulated environments, and compare todiversity-based baselines. The findings suggest that our approach yields a moreefficient and robust policy search process.
Conference paperShen M, Clark A, Rojas N, 2020,
Robotic exoskeletons have been a focal point of research due to an ever-increasing ageing population, longer life expectancy, and a desire to further improve the existing capabilities of humans. However, their effectiveness is often limited, with strong rigid structures poorly interfacing with humans and soft flexible mechanisms providing limited forces. In this paper, a scalable variable stiffness revolute joint is proposed to overcome this problem. By using layer jamming, the joint has the ability to stiffen or soften for different use cases. A theoretical and experimental study of maximum stiffness with size was conducted to determine the suitability and scalablity of this technology. Three sizes (50 mm, 37.5 mm, 25 mm diameter) of the joint were developed and evaluated. Results indicate that this technology is most suitable for use in human fingers, as the prototypes demonstrate a sufficient torque (0.054 Nm) to support finger movement.
Conference paperWang J, Lu Q, Clark A, et al., 2020,
A passively compliant idler mechanism for underactuated dexterous grippers with dynamic tendon routing, Towards Autonomous Robotic Systems Conference (TAROS ) 2020, Publisher: Springer Verlag, Pages: 25-36, ISSN: 0302-9743
In the field of robotic hands, tendon actuation is one of the most common ways to control self-adaptive underactuated fingers thanks to its compact size. Either differential or direct drive mechanisms are usually used in these systems to perform synchronised grasping using a single actuator. However, synchronisation problems arise in underactuated grippers whose position of proximal joints varies with time to perform manipulation operations, as this results in a tendon-driven system with dynamic anchor pulleys. This paper introduces a novel passively compliant idler mechanism to avoid unsynchronisation in grippers with a dynamic multi-tendon routing system, such that adequate grasping contact forces are kept under changes in the proximal joints’ positions. A re-configurable palm underactuated dexterous gripper is used as a case study, with the performance of the proposed compliant idler system being evaluated and compared through a contact force analysis during rotation and translation in-hand manipulation tasks. Experiment results clearly demonstrate the ability of the mechanism to synchronise a dynamic tendon routing gripper. A video summarising experiments and findings can be found at https://imperialcollegelondon.box.com/s/hk58688q2hjnu8dhw7uskr7vi9tqr9r5.
Journal articleFischer T, Demiris Y, 2020,
Humans are inherently social beings that benefit from their perceptional capability to embody another point of view, typically referred to as perspective-taking. Perspective-taking is an essential feature in our daily interactions and is pivotal for human development. However, much remains unknown about the precise mechanisms that underlie perspective-taking. Here we show that formalizing perspective-taking in a computational model can detail the embodied mechanisms employed by humans in perspective-taking. The model's main building block is a set of action primitives that are passed through a forward model. The model employs a process that selects a subset of action primitives to be passed through the forward model to reduce the response time. The model demonstrates results that mimic those captured by human data, including (i) response times differences caused by the angular disparity between the perspective-taker and the other agent, (ii) the impact of task-irrelevant body posture variations in perspective-taking, and (iii) differences in the perspective-taking strategy between individuals. Our results provide support for the hypothesis that perspective-taking is a mental simulation of the physical movements that are required to match another person's visual viewpoint. Furthermore, the model provides several testable predictions, including the prediction that forced early responses lead to an egocentric bias and that a selection process introduces dependencies between two consecutive trials. Our results indicate potential links between perspective-taking and other essential perceptional and cognitive mechanisms, such as active vision and autobiographical memories.
This data is extracted from the Web of Science and reproduced under a licence from Thomson Reuters. You may not copy or re-distribute this data in whole or in part without the written consent of the Science business of Thomson Reuters.