README.md

Environment

Available robots

• Panda, the Franka Panda Emika robot

Goal-conditioned tasks

• Reach
• Push
• Slide
• PickAndPlace
• Stack{2..4}

Language-conditioned tasks

• NLReach
• NLPush
• NLGrasp
• NLLift

Schema

The following schema is used to specify the environments for gym.make():

{Robot Name}{Task Name}{Number of Objects}[Mode][Observation Type][Hindsight Instructions][Action Correction].

Parameters in {} are required and [] are optional.

Number of objects

• 2
• 3

Mode

• (square with 3 colors)
• Color (square with 9 colors)
• Shape (3 shapes with 3 colors)
• ColorShape (3 shapes with 9 colors)
• Weight (square with 2 different weight properties)
• Size (square in 2 different sizes)
• SizeShape (3 shapes in 2 different sizes)
• WeightShape (3 shapes with 2 different weight properties)

Observation type

• (default): empty string to use world state (incl. Kinematic information) as observation
• PixelStatic (static RGB image of the scene)
• PixelEgo (egocentric RGB image)

Hindsight Instructions

• empty string
• HI to add hindsight instructions once to info dict of gym.step() when interaction with wrong object is detected

Action Correction

• empty string
• AR to enable action correction
• ARN to enable action correction with negations
• ARD to enable action correction with a small delay of up to 20 time steps
• ARND to enable action correction with negations and a small delay of up to 20 time steps