Sign into an existing account

The basic idea behind imitation learning is implicitly giving a robotic entity prior information about the world by mimicking human behaviour. Through such techniques, which can be likened to behaviour cloning or inverse reinforcement learning, it is possible to increase the efficiency of robots and their interaction capacities with humans, while also decreasing the cost involved in learning new skills.

Inspired by the idea that a robotic manipulator can imitate any visually-demonstrated behaviour of arbitrary complexity, a DeepMind research team has proposed manipulation-independent representations (MIR) that can support successful imitation of behaviours demonstrated by previously unseen manipulator morphologies using only visual observations.

image.png
The researchers first explain how to imitate unconstrained manipulation trajectories executed by previously unknown manipulators, specifically how to learn pixel-based representations that produce the right information for a cross-embodiment trajectory tracking agent. The proposed imitation method has two main phases: learning an MIR space, and cross-embodiment visual imitation through reinforcement learning using the pretrained MIR space.

In the evaluations, MIR achieved the best performance across all test domains, significantly boosting Jaco Hand performance on stacking success and perfectly imitating the simulated Jaco Hand and Invisible Arm on lifting, with a 100 percent score.

Overall, the study revealed the important considerations in learning manipulator and task-independent representations for imitation, and validated the proposed MIR’s characteristics for successful cross embodiment visual imitation.

The paper Manipulator-Independent Representations for Visual Imitation is on arXiv.

https://arxiv.org/pdf/2103.09016.pdf