vrgripper

VRGripper Environment Models

Contains code for training models in the VRGripper environment from "Watch, Try, Learn: Meta-Learning from Demonstrations and Rewards."

Includes the models used in the Watch, Try, Learn (WTL) gripping experiments.

Links

• Project Website
• Paper Preprint

Authors

Allan Zhou¹, Eric Jang¹, Daniel Kappler², Alex Herzog², Mohi Khansari², Paul Wohlhart², Yunfei Bai², Mrinal Kalakrishnan², Sergey Levine^1,3, Chelsea Finn¹

¹ Google Brain, ²X, ³UC Berkeley

Training the WTL gripping experiment models.

WTL experiment models are located in vrgripper_env_wtl_models.py. Data is not included in this repository, so you will have to provide your own training/eval datasets. Training is configured by the following gin configs:

• configs/run_train_wtl_statespace_trial.gin: Train a trial policy on state-space observations.
• configs/run_train_wtl_statespace_retrial.gin: Train a retrial policy on state-space observations.
• configs/run_train_wtl_vision_trial.gin: Train a trial policy on image observations.
• configs/run_train_wtl_vision_retrial.gin: Train a retrial policy on image observations.