Learning Intention Aware Online Adaptation of Movement Primitives

Abstract

In order to operate close to non-experts, future robots require both an intuitive form of instruction accessible to laymen and the ability to react appropriately to a human co-worker. Instruction by imitation learning with probabilistic movement primitives (ProMPs) allows capturing tasks by learning robot trajectories from demonstrations including the motion variability. However, appropriate responses to human co-workers during the execution of the learned movements are crucial for fluent task execution, perceived safety, and subjective comfort. To facilitate such appropriate responsive behaviors in human-robot interaction, the robot needs to be able to react to its human workspace co-inhabitant online during the execution of the ProMPs. Thus, we learn a goal-based intention prediction model from human motions. Using this probabilistic model, we introduce intention-aware online adaptation to ProMPs. We compare two different novel approaches: First, online spatial deformation, which avoids collisions by changing the shape of the ProMP trajectories dynamically during execution while staying close to the demonstrated motions and second, online temporal scaling, which adapts the velocity profile of a ProMP to avoid time-dependent collisions. We evaluate both approaches in experiments with non-expert users. The subjects reported a higher level of perceived safety and felt less disturbed during intention aware adaptation, in particular during spatial deformation, compared to non-adaptive behavior of the robot.

User intent estimation during robot learning using physical human robot interaction primitives

“…human robot gestures (Ben Amor et al 2014), collaborative object covering (Cui et al 2019), and hand shaking (Campbell et al 2019). The probabilistic model embeds the user intent in their outputs, making it suitable to utilize probabilistic operators to adapt the model for goal and trajectory adaptation (Bajcsy et al 2017;Koert et al 2019), sequential intent estimation (Matsubara et al 2015), and stiffness adaptation (Rozo et al 2016).…”

Learning to Pick at Non-Zero-Velocity from Interactive Demonstrations

“…Different works explore speed adaptation during trajectory execution using different function approximators. One approach involves altering the phase rate of probabilistic movement primitives (ProMPs) [11], [12], whereas another proposes the use of a modified version of Dynamical Movement Primitives (DMPs) in which the speed is altered through an additional phase-dependent temporal scaling factor [5], [13]. While the focus of these existing works is in combining imitation learning with reinforcement learning, our approach combines imitation learning and human interactive corrections [14].…”

Perceived Safety in Physical Human Robot Interaction -- A Survey

“…Also, the concepts of valence and arousal were employed in [21,23,24], as part of a general framework used to describe the emotional experience of the participants. In addition to the previous categorization, papers [17, 20, 21, 23-26, 28, 29, 31-33, 35, 36, 38-40] focused on finding a connection between perceived safety, robot speed, and relative distance between human and robot, with [27,30,40] having an explicit focus on motion prediction. The human attitude toward robot position and approach direction was studied in [18,25].…”

“…Koert et al [40] focused on generating the robot motion via imitation learning with probabilistic movement primitives, by proposing two methods (based on spatial deformation and temporal scaling) for real-time human-aware motion adaptation. The main aim of the work was to guarantee perceived safety and comfort, using a goal-based intention prediction model learnt from human motions.…”