top of page

Meta-Reinforcement Learning via Language Instructions

Zhenshan Bing, Alexander Koch, Xiangtong Yao, Kai Huang, and Alois Knoll

The picture captured by a RealSense D435i camera.

Push, Pick-place tasks

Push task with different goals.

Pick-place task with different goals.

According to the task setting of the Meta-World benchmark, the agent successfully solves the task, if the distance between the object and the task goal is less than 5cm. Therefore, in these videos, there is a short distance between the object and the task goal, but the distance is less than 5cm.

We adapt Panda robot to Meta-world environment. For more information, please visit 

Real-world experiments

scene.jpg

Real-world experiment scenario. We use Intel RealSense D435i cameras to locate the red object.

scene2.jpg

Supplementary figures

ml10_v2.jpg

Meta-World ML10 V2 training progress with language instructions. The shaded regions indicate one standard deviation over three training runs.

Language instructions for all Meta-World tasks.

bottom of page