An interactive motion planning framework that can learn from experience
Introduction
The verification of accessibility in assembly and disassembly is an essential issue for the product design. However, the accessibility verification presents challenges to both humans and computers. For humans, the manual manipulation of 3D object in a complex virtual environment is extremely difficult [1]. For computers, the accessibility verification can be defined as computing a collision-free path for an object between a free location and a pre-designed location. In the last decade, the sampling based motion planners have been successfully used to solve the motion planning problems [2]. However, the performance of sampling based planning algorithms may degrade if the collision-free space contains narrow passages [3], i.e. small regions whose removal or perturbation can change the connectivity of the collision-free space [4]. In particular, the narrow passages commonly exist in the virtual assembly/disassembly tasks. Moreover, if no collision-free path exists, the sampling based algorithms may run for a long time before verifying that [5]. Different heuristics have been created to bias the automatic planners. Decomposition based planners decompose the configuration space into different levels of cells and get an approximate corridor with algorithm, then explore this corridor to find a collision-free path [6], [7]. Obstacle based planners either plan motion compliant to the -contact space (-contact space is the subset of configuration space, which consists of the configurations when the robot touches one or more obstacles without any penetration [4]) with geometric formulations [8], [9], or increase the sampling around the obstacle [10], [11] and in the -contact space [12]. In [13], the vectors generated on the triangulated meshes of CAD models are used to guide the extension of basic RRT planner (Rapidly-exploring Random Trees). Retraction based planners attempt to improve the sampling in narrow passages through retracting the in-collision configuration to the nearest configuration in the -contact space [4], [14]. Despite some improvements, the narrow passage is still a major issue for the sampling based planners. The weak consciousness to the local changes of geometric space results in the above difficulties.
Moreover, designers need to test and verify the geometric constraints to the articulate system of a human skeleton. A direct method is adding a digital human mannequin into the reasoning loop of an automatic method [1]. The quality of the generated motion of the mannequin becomes a major issue, which mainly concerns reducing the redundant degrees of freedom and making the motion look more natural. Besides, a key challenge for the designers is to analyze end-users’ responses in the design process and to promote product innovation, but users’ preferences cannot be precisely formulated [15]. Hence, the optimization of the preference of users is today extremely difficult for an automatic method, and the integration with the direct participation of designers is an essential stage [16], [17].
A solution to the above challenges is to make humans interact with an automatic motion planner. By relaxing the collision constraints, it is easier for a human to find a motion solution. Interactive motion planners attempt to integrate humans with the automatic path planners. The method in [18] enables users with an haptic device to visually specify some critical via-configurations or sub-goals in the workspace, which guide the exploration of a sampling based planner. However, in the practice, some via-configurations in complex environments are unlikely to be found by humans, or unreachable. In [6], [19], a preliminary discretization of configuration space is carried out by a decomposition algorithm, and then a corridor between the initial configuration and the goal configuration is found. In this corridor, the harmonic functions are formulated to provide force feedback and a sampling based planner searches a feasible path. Although these kind of methods successfully accelerate the automatic planner by reducing the search space, humans impose very limited impact on the planning process. The method proposed in [20] is successfully applied to the interactive motion planning of a steerable needle, but not applicable to a non-convex object. In [21], the method simply allows the simultaneous cooperation between users and automatic algorithms. But humans and automatic planners cooperate in a relatively independent manner. In the “narrow passages”, both of them may fall into a trap and cannot guide each other efficiently.
Learning experience to serve motion planning in new environments has been studied by various works. Recent research focuses on exploiting the accumulated data in the prior planning results. In [22], the approach proposes to learn from the prior instances of collision queries and to improve the sampling based motion planners in avoidance of the exact collision checking. The “retrieve and repair” idea of [23] is similar to ours. However, their retrieval process based on the amount of constraints violated by a path is implemented in a robot’s configuration space, which limits their adaptability to quite different environments. In [24], a more robust and adaptive motion learning method is proposed to map from the previously optimized trajectories in old situations to the good trajectories in new situations. Although similar to our idea of retrieving paths based on the similarity of workspace, their method is limited to an offline constructed motion library composed of optimized motion trajectories. In contrast, the concept of online motion learning based on the topological similarity between scenarios has been proposed in [25], however, no concrete or workable method was developed by the authors. Given the workspace ( in 2D; in 3D) and its subset taken up by the obstacles , the scenario for a moving object is the subset of the workspace excluding the obstacles, which is noted as . In this paper, the relationship between the topology of scenarios and the motion of objects is discovered, and based on it a novel scenario retrieval and motion reuse strategy is proposed.
In this paper, a novel framework for the interactive motion learning and generation is proposed. The learning is dynamic: the motion library can constantly evolve from an empty one to a very large one, and it is maintained by a balanced hierarchical structure. More detailed contributions include:
- •
The use of the maximal sphere sequence to represent the scenario. The centerline and the radius change of the sphere sequence can respectively describe the topological structure and the volume change of the scenario. The method of the real time generation of the maximal sphere sequence is also provided.
- •
A novel motion learning framework based on computing the similarity of the topological structure and the volume change of scenarios.
- •
The great acceleration of the motion planning process, especially in narrow passages, thanks to the adequate learning and motion generation strategy.
- •
Simplified ways of interacting with a computer by releasing humans from the rotational manipulations, which make it possible to use non-haptic devices in a complex assembly/disassembly task.
We have applied our algorithms and demonstrated its effect in a challenging case in 2D. The rest of the paper is organized in the following manner. Section 2 briefly surveys the related tools in other domains. Section 3 is dedicated to the detailed explanation of our framework. Simulation results are presented and analyzed in the Section 4. The limitations and extensions are discussed in the Section 5.
Section snippets
Related tools from other domains
The recent developments in the fields of motion primitive, medial axis, curve skeleton and curve matching are the basis of our motion learning and generation framework.
Interactive motion learning and generation
In this section, a hypothesis is proposed to describe the relationship between the motion of an object and its corresponding scenario. Then an overview of our motion learning and generation framework is demonstrated prior to the detailed explanations of each functional module.
Implementation and results
In the following section, our motion planning framework will be applied to a non-convex object moving in a cluttered 2D environment which contains the randomly shaped and located non-convex obstacles. The experimental results demonstrate the efficiency of our method. Moreover, after adequate learning, the dependence on the haptic devices can be greatly reduced in the interaction process by releasing users from the rotation manipulations.
All the experimental results are performed on the software
Conclusion and extension
Based on the hypothesis that the motion of an object has a close relationship with the topological structure and the volume size of the scenario, we proposed our interactive motion learning and generation framework, and then applied it to a challenging motion planning problem in 2D. For the first time, the topological similarity between scenarios is used to explore the patterns of motion in configuration space. The experimental results verified our hypothesis. Due to the highly efficient
References (49)
- et al.
Improving assembly precedence constraint generation by utilizing motion planning and part interaction clusters
Comput-Aided Des
(2013) - et al.
Affective design of products using an audio-based protocol: application to eyeglass frame
Int J Ind Ergon
(2014) - et al.
Integrated and interactive method for solving layout optimization problems
Expert Syst Appl
(2013) - et al.
Haptic-aided robot path planning based on virtual tele-operation
Robot Comput-Integr Manuf
(2009) - et al.
Curve skeleton extraction by coupled graph contraction and surface clustering
Graph Models
(2013) - et al.
Curve matching for open 2D curves
Pattern Recognit Lett
(2009) - et al.
Solving inverse configuration space problems by adaptive sampling
Comput-Aided Des
(2013) - et al.
Generalized penetration depth computation
Comput-Aided Des
(2007) - Ferré E, Laumond J-P, Arechavaleta G, Estevès C. Progresses in assembly path planning. In: Proceedings of international...
- Kuffner JJ, LaValle SM. RRT-connect: an efficient approach to single-query path planning. In: Proceedings of...