Multi-controller multi-objective locomotion planning for legged robots

Abstract

Different legged robot locomotion controllers have different advantages and disadvantages, from speed of motion to energy, computational speed, safety and others. In this paper we propose a method for planning locomotion with multiple controllers and sub-planners, explicitly considering the multiobjective nature of the problem. We propose a parameter-free method that plans in the space of body motion and controller choice, using utopian and lexicographic cost aggregation functions. We empirically analyze the behavior of the method in terms of planning success rates, Pareto-optimality and anytime behavior in cost space. We show that our method is faster than pure footstep planning methods both in computation (2x) and mission time (1.4x), it is safer than pure dynamic-walking methods, and achieves desirable Pareto-optimal solutions (up to 8x) faster than fairly-tuned traditional weighted-sum methods. Our conclusions are drawn from a combination of planning, physics simulation, and real robot experiments.

GaitMesh: Controller-Aware Navigation Meshes for Long-Range Legged Locomotion Planning in Multi-Layered Environments

“…an agent could swim over a river or travel to the closest bridge). This functionality is also close to the reality of legged robot locomotion-which is characterized by a large set of possible gaits [24] and controllers specialized for different kinds of terrain [1]. In this paper we use navigation meshes with controller-choice annotations to make long-term plans aware of the real cost of traversing different regions, and hence obtain paths of low global cost.…”

“…These are also popular for legged robot locomotion planning. For example, [10] uses search on occupancy grids and [11], [1] on heightmaps.…”

“…Our work targets long-term and large-scale robot deployments, for quadruped robots in particular. We specifically focus on building a system for fast path-finding in large-scale, potentially multi-level environments such as industrial oil rigs and other hard-to-access sites [1]. Except for PRMs, the map representations we have mentioned are either single-floor or manually designed-which could lead to imprecise maps or large manual work requirements.…”

“…We use GridMap for heightmap computations [9]. We use sensor-based maps instead of the NavMesh at this stage in order to reflect 1 The repository will be made public upon acceptance.…”

“…However, it comes at the cost of reasoning about terrain features and choosing the appropriate gait to use for each particular area. Even though the choice of gaitcontroller has been included in recent locomotion planning methods [1], it is still not clear which map representations and planning methods are most suitable to the task-especially for large-scale environments.…”