One-Shot Robust Imitation Learning for Long-Horizon Visuomotor Tasks from Unsegmented Demonstrations

TL;DR

Dynamical movement primitives and meta-learning are exploited to provide a new framework for imitation learning called Meta-Imitation Learning with Adaptive Dynamical Primitives (MiLa), which allows for learning unsegmented long-horizon demonstrations and adapting to unseen tasks with a single demonstration.

Abstract

In contrast to single-skill tasks, long-horizon tasks play a crucial role in our daily life, e.g., a pouring task requires a proper concatenation of reaching, grasping and pouring subtasks. As an efficient solution for transferring human skills to robots, imitation learning has achieved great progress over the last two decades. However, when learning long-horizon visuomotor skills, imitation learning often demands a large amount of semantically segmented demonstrations. Moreover, the performance of imitation learning could be susceptible to external perturbation and visual occlusion. In this paper, we exploit dynamical movement primitives and meta-learning to provide a new framework for imitation learning, called Meta-Imitation Learning with Adaptive Dynamical Primitives (MiLa). MiLa allows for learning unsegmented long-horizon demonstrations and adapting to unseen tasks with a single demonstration. MiLa can also resist external disturbances and visual occlusion during task execution. Real-world robotic experiments demonstrate the superiority of MiLa, irrespective of visual occlusion and random perturbations on robots.

Figures (6)

• Figure 1: An overview of the MiLa framework. After learning from $\boldsymbol{\mathrm{d}}^{\text{tr}}$ (indicated by yellow arrows), MiLa acquires the capability to adapt to new tasks $\boldsymbol{\mathrm{d}}^{\text{val}}$ (see blue arrows). Instead of predicting robot actions as per visual inputs at each time step, MiLa predicts task parameters for each subtask and a set of dynamical movement primitives are employed to generate robot trajectories across different subtasks.
• Figure 2: The experimental setup as well as objects for training and testing in long-horizon tasks.
• Figure 3: Demonstrations utilized to establish the skill repertoire $\{\rho^c\}_{c=1}^3$. Dynamical motion primitives for the reaching, placing, and pushing skills are learned from demonstrations depicted in (a)--(c), respectively. Additionally, we collect 5 demonstrations for each skill to model the intrinsic skill variability, as shown in (d)--(f).
• Figure 4: Snapshot of the long-horizon task evaluations. First row shows the kinesthetic teaching of the reaching-placing-pushing task. Second and third rows correspond to the evaluations of MiLa and MiLa-NoWeight, respectively. Fourth row illustrates an evaluation of MAML. Fifth and sixth rows present the success and fail cases using GCBC, respectively.
• Figure 5: Snapshot of the long-horizon task in the presence of visual occlusion. First and second rows correspond to evaluations using MiLa, with the depth camera's perspective and the user's view, respectively. Similarly, third and fourth rows correspond to evaluations using MAML. Fifth and sixth rows are evaluations with GCBC.