Active Task Randomization: Learning Robust Skills via Unsupervised Generation of Diverse and Feasible Tasks
Kuan Fang*, Toki Migimatsu*, Ajay Mandlekar, Li Fei-Fei, Jeannette Bohg
(* indicates equal contributions)
Abstract
Real-world sequential manipulation tasks require robots to chain several manipulation skills together across a broad range of situations. Learning such general-purpose skills can require large-scale, diverse training data, which can be labor-intensive and non-trivial to collect and curate. In this work, we introduce Active Task Randomization (ATR), an approach that learns a repertoire of visuomotor skills for sequential manipulation by automatically creating feasible and novel tasks in simulation. During training, our approach procedurally generates tasks using a graph-based task parameterization. To adaptively estimate feasibility and novelty of sampled tasks, we develop a relational neural network that projects each task parameter into a compact embedding. We demonstrate that our approach can automatically create suitable tasks for efficiently training the skill policies to handle diverse scenarios. We evaluate our method on simulated and real-world sequential manipulation tasks by composing the learned skills using a task planner. Compared to baseline methods, the skills learned using our approach consistently achieves superior success rates.
Examples of Proposed Tasks
Tasks of various object types, scales, symbolic relationships, and camera configurations are proposed in simulation. Most of the proposed tasks are feasible and can be solved by the planned actions. The target objects are denoted by the same colors as they appear in the simulated environment.
Real-World Sequential Manipulation
ATR trains visuomotor skills by actively generating tasks that are challenging and unfamiliar for the policies. As a result, the skill policies gain broader experience through ATR than through other baselines and are able to generalize well to unseen tasks with unseen objects. The videos below demonstrate how the trained skills can be used to solve sequential manipulation tasks in the real world.
(The loading of the videos below could take several seconds. Thank you for your patience!)
Task A
place-onto(sugar, table)
place-onto(yogurt, cream)
place-onto(milk, container)
Task B
push-under(container, rack)
place-nextto(milk, rack)
place-onto(salt, rack)
Task C
pull-with(icecream, hook)
place-nextto(icecream, rack)
push-under(icecream, rack)
place-onto(pasta, rack)
