The main ingredients for legged locomotion are planning, control, perception and state estimation. in our locomotion framework, we envisioned different layers, according to the quality of the perception feedback available. At the bottom layer we have blind reactive strategies, with basic terrain adaptation mechanisms and reflex strategies.
Motion primitives are triggered to override the planner in situations where the robot could be damaged (e.g. stumbling). We also implemented several gaits (crawl, trot and bounding) that allow our quadruped robots traverse challenging terrains. However, when the complexity of the terrain to be negotiated increases, or when the execution of the requested task involves highly dynamic motions, optimization-based planning strategies are preferable. Therefore, another branch of studies is aiming to reduce the complexity of the optimization, trading off accuracy with computational efficiency, in order to be able to carry out (online) optimization of trajectories during locomotion.
Planner-free Whole-Body Locomotion Framework for Quadruped Robots
This work proposes a hierarchical whole-body framework able to generate different kinds of gaits, ranging from very dynamic gaits such as the trot, to more static gaits like the crawl. The framework is composed of a priority-based whole-body controller that works in synergy with a walking pattern generator and does not require any higher level planning for the CoM trajectories (i.e. is planner-free), nor state estimation.
The approach introduces a smart use of the postural task to remove indeterminacy and achieves consistency between feet and base motion while aiming to keep a well behaved kinematic configuration. In particular, the postural task exploits the redundancy of the higher priority tasks to keep the robot close to a preferable nominal configuration. This generates a connection between the motion of the trunk and the location of the contacts. Therefore, the postural task acts as a set of ``elastic linkages'' and determines the linear motion of the base, aligning it with the feet.
The effectiveness of their locomotion framework was demonstrated on different quadruped platforms such as the Hydraulically-actuated Quadruped (HyQ) and ANYmal.
STANCE: Locomotion Adaptation over Soft TerrainWhole-body Control (WBC) can achieve multiple locomotion tasks while respecting the robot’s dynamics. However, most of WBC frameworks fail to generalize beyond rigid terrains. Over soft terrain, the stability and performance of the legged robot deteriorate if the WBC is not accounting for the introduced uncertainty due to the soft contact interaction. To deal with this, we propose a novel soft terrain adaptation algorithm called STANCE: Soft Terrain Adaptation and Compliance Estimation. STANCE consists of a WBC that exploits the knowledge of the terrain to generate an optimal solution that is contact consistent and an online terrain compliance estimator that provides the WBC with terrain knowledge.
Feasible Wrench Polytope (FWP) and the Feasible RegionSimplified dynamic models can be used in mathematical programming to plan optimal robot trajectories (Center of Mas positions and velocities, orientation and angular speed, feet positions and contact forces). All common reduced models, however, completely omit the joint-space torque limits of the system, that are one of the most common causes of failure of a locomotion task. In this section, we introduce two articles that propose to project the high dimensional joint-torque limits into lower dimensional spaces (either 6D or 2D).
Whole-body Control for Quadruped RobotsIn dynamic locomotion, it is essential to consider the robot’s dynamics, actuation limits and interaction with the environment. To do so, we exploit optimization techniques in locomotion control to formulate a Whole-Body Control (WBC). WBC is a motion tracking controller capable of achieving multiple locomotion tasks while respecting the robot’s behavior.
Visual Foothold Adaptations Using Machine LearningWith a better understanding of the terrain, legged robots become substantially more versatile. Visual knowledge of the terrain is key to avoid detrimental scenarios such as falling or slipping. However processing visual information to assess the terrain is in general computationally costly. This prevents from having fast and precise reactions needed to promptly recover from unexpected situations, e.g., slippage and surface collapse. To reduce such computational load, we use machine learning-based strategies, to effectively negotiate obstacles and traverse rough terrain in a continuous fashion.
Omni-directional Bounding GaitWe present the steps that allowed us to realize real outdoor experiments of HyQ bounding at different speeds and performing omni-directional maneuvers. The strategy is composed of two parts: the first one is an offline optimization that finds a stable periodic limit cycle which represents the base-line bounding gait; the second part is a speed controller that adjusts online the main gait parameters based on the high-level speed commands coming from the external operator. In the tests HyQ reached a forward speed of 2.5m/s, lateral speed of 1m/s and angular speed of 50 deg/s in simulation and, respectively, 1 m/s, 0.5m/s and 30deg/s on the hardware experiments.
Reactive Controller FrameworkWhile there is a lot of work addressing single aspects of the overall locomotion gait planning and control problem, solutions that take all these elements together in a systematic and coherent fashion are rare. In this contribution we present a reactive gait generation and control framework for a quadruped robot that has the following main goals:
- Creation of stable omni-directional periodic gait
- Robustness against disturbances from uneven ground and external forces on the trunk
- Foot slip avoidance
- Reduction of impact forces at the feet
To achieve these goals we make extensive use of kinematic and dynamic models of our hydraulic quadruped robot HyQ and exploit the high performance torque-control available at all the joints. The resulting control framework exhibits the following features (Barasuol, 2013):
- Avoid trajectories that would penetrate the ground (avoid high ground reaction forces – GRF);
- Avoid weak contact or loss of contact (avoid slippage);
- Avoid undesired leg internal forces (avoid slippage and waste of energy);
- Reduce disturbances between joint position and trunk attitude controllers;
- Reduce disturbances at the trunk due to poor state estimation (avoid excessive GRF);
- Increase the locomotion robustness with respect to unexpected terrain irregularities (avoid excessive GRF).
Crawl on Rough TerrainWe present a heuristic-based planning approach for rough terrain locomotion. The terrain adaptation and the mitigation of tracking errors is achieved through a 1-step online replanning strategy, which can work either blindly or with visual feedback. We show real experiments where HyQ is able to successfully negotiate different types of terrain templates (ramps, debris, stairs, steps), some of them with a significant roughness. The size of the stones are up to 12 cm (diameter) that is about 26% the retractable leg length. We also applied this strategy, with little variations, to the task of climbing up and down industrial-size stairs, also performing 90deg turns while climbing the stairs.
Momentum Based Disturbance ObserverA significant source of errors can come from unmodeled disturbances, such as an external push. If a proper identification is carried out offline, the updated model can then used for a more accurate inversion of the dynamics. Even though this approach is effective to detect constant payload changes, it is not convenient to estimate time-varying unknown external forces, which might change both in direction and intensity. Indeed, this kind of disturbances can dramatically increase the tracking errors and jeopardize the locomotion, unless they are compensated online. We developed a Momentum Based Disturbance Observer, capable of estimating both linear and angular components of an external disturbance wrench of variable amplitude applied on the robot during the execution of a walk. The identified wrench is compensated during the locomotion, improving the tracking accuracy and locomotion stability of HyQ that is climbing a ramp while pulling a wheel barrow (27kg load).
Simultaneous Contact and Motion PlanningTraditional motion planning approaches for multi-legged locomotion divide the problem into several stages, such as contact search and trajectory generation. However, reasoning about contacts and motions simultaneously is crucial for the generation of complex whole-body behaviors. We propose a mixed-integer convex formulation to plan simultaneously contact locations, gait transitions and motion, in a computationally efficient fashion.our approach is not limited to flat terrain nor to a pre-specified gait sequence.We experimentally validated the approach on the HyQ robot by traversing different challenging terrains, where non-convexity and flat terrain assumptions might lead to sub-optimal or unstable plans.
Online Payload IdentificationThe identification of inertial parameters is crucial to achieve high-performance model-based control of legged robots. The inertial parameters of the legs are typically not altered during expeditions and therefore are best identified offline. On the other hand, the trunk parameters depend on the modules mounted on the robot. We proposed multiple original techniques to identify online the parameters of the trunk of the robot. This is meant to improve the accuracy of the dynamic model and ensure locomotion stability, whenever a change in the load is detected. We show experiments of HyQ walking on flat and rough terrain while we add a weight on its trunk. The robot detects the payload change, performs the online identification and continues to walk with the updated parameters.
The Height ReflexThe height reflex is a motion generation strategy for all the robot’s feet. It redistributes the swing motion onto the stance legs, with the final effect of “lowering” the trunk to assist the foothold searching motion. The height reflex is useful when the robot is facing consid- erable changes in the terrain elevation (e.g., stepping down from a high platform). In such situations, the swing leg can lose mobility, causing issues during the subsequent steps (e.g., walking with excessively stretched legs). For this reasons the height reflex is most likely to be activated when stepping down.
The Step ReflexThe step reflex is a local elevator recovery strategy, triggered in cases of frontal impacts with an obstacle during the swing up phase. This reflex is key in cases of visual deprivation (e.g., smoky areas or thick vegetation): it allows the robot to overcome an obstacle and establish a stable foothold at the same time, without stumbling. If a frontal impact is detected in the swing down phase, the reflex is not triggered immediately, but it is scheduled for the beginning of the next swing phase. Even if the swing leg stumbles against the step, the robot is still able to climb the stairs, thanks to the step reflex.
The Chimney ClimbResearch into legged robotics is primarily motivated by the prospects of building machines that are able to navigate in challenging and complex environments that are predominantly non-flat. In this context, control of contact forces is fundamental to ensure stable contacts and equilibrium of the robot. In this paper we propose a planning/control framework for quasi-static walking of quadrupedal robots, implemented for a demanding application in which regulation of ground reaction forces is crucial. We show experimental results where HyQ is able to walk inside two high-slope (50◦) V-shaped walls; an achievement that to the authors’ best knowledge has never been pre-sented before. The robot distributes its weight among the stance legs so as to optimize user-defined criteria.
De-coupled Planned walking over rough terrain
- Α framework for quadrupedal locomotion over highly challenging terrain where the choice of appropriate footholds is crucial for the success of the behavior.
- we developed a path planning approach that evaluates the geometry of the environment, computes desired body trajectories and selects locally optimal footholds along the planned path.
- We exploit the active compliance of our system to smoothly interact with the environment, even when this is inaccurately perceived or dynamically changing, and update the planned path on-the-fly.
- We leverage the full set of benefits that a high performance torque controlled quadruped robot can provide and demonstrate the flexibility and robustness of our approach on a set of experimental trials of increasing difficulty.
- A. Winkler, C. Mastalli, I. Havoutis, M. Focchi, D. G. Caldwell, C. Semini, Planning and Execution of Dynamic Whole-Body Locomotion for a Hydraulic Quadruped on Challenging Terrain, IEEE International Conference on Robotics and Automation (ICRA), 2015 [full article]
- C. Mastalli, A. Winkler, I. Havoutis, D. G. Caldwell, C. Semini, On-line and On-board Planning and Perception for Quadrupedal Locomotion, 2015 IEEE International Conference on Technologies for Practical Robot Applications (TEPRA), 2015 [full article]
- A. Winkler, I. Havoutis, S. Bazeille, J. Ortiz, M. Focchi, R. Dillmann, D. G. Caldwell, C. Semini, "Path Planning with Force-Based Foothold Adaptation and Virtual Model Control for Torque Controlled Quadruped Robots," IEEE International Conference on Robotics and Automation (ICRA), 2014. [full article]